* changes v7.1.2 -> v7.1.x

##
## Release Highlights
##

This release expands candidate generation, improves large-scale operations and removes several
longstanding bottlenecks in hash management, device setup and distributed work.

1. Plugin-based candidate generation with attack-mode 8: "feeds"

Attack-mode 8 brings external candidate generators into hashcat as feed plugins. Unlike a generator
connected through stdin, a feed runs one instance per device and participates in keyspace
calculation, distribution, restore, status and debug output. Feeds can be written in C or Rust and
selected by name.

The included feeds are "wordlist", "stdin", "random", "rust_random", "association", "pcfg",
"table", "hybrid" and "mask". Wordlist and stdin processing now use the feed interface. The
wordlist feed supplies the dictionaries used by attack-modes 0, 1, 6, 7 and 12, and the one word
per hash form of attack-mode 9. With rules, the hybrid and mask feeds generate candidates for
attack-modes 1, 3, 6, 7 and 12 on the host.

The wordlist feed lays every source it receives end to end in one keyspace, including multiple
wordlists and directories. Mask files and increment lengths likewise form one queue. A position
addresses the whole feed or mask queue rather than one source, so "--skip", "--limit" and
"--keyspace" divide that set and a device can start at its own share instead of reading and
discarding every earlier line. The seek databases described below make that start cheap. Dynamic
dictionary queues and association phases remain separate rounds, as described under migration.

The feed interface also introduces backend-controlled GPU execution for candidate generators. A
feed can compile and launch a kernel of its own without directly managing the compute backend. The
PCFG and table feeds use this to expand base words into candidates in device memory, avoiding a
transfer across the host-to-device bus. They use device amplification for most fast hash modes and
fall back to the host engine where needed. The device engine derives its own loop bounds, so it
rejects "-u". Rules with "-r" or "-g" also select host generation.

See "docs/hashcat-generic-attack-mode.md" and
"docs/hashcat-generic-attack-mode-development-guide.md".

2. Probability-ordered PCFG and targeted hint attacks with attack-mode 4

This new attack-mode generates candidates from a trained probabilistic context-free grammar.
Hashcat quantizes component probabilities into cost groups whose sizes are known in advance. A
component costs "-log2(p) * scale" rounded to a whole number, so a larger probability is a smaller
cost and the number of candidates at each cost can be counted without producing them. This makes a
PCFG attack compatible with the options "--keyspace", "--skip", "--limit", "--restore",
multi-device distribution and the brain.

The command "-a 4 hashes.txt" selects the included "default-passwords" ruleset. Rulesets can be
selected by name or path, combined with weights, and stored as directories or ".tar.xz" archives.

The second included ruleset "hints" is a targeted PCFG attack, and it is the part of this release
worth looking at first. It removes the words learned during training but retains the surrounding
password patterns. The operator supplies facts about one target through "hintwords" or "hintfile",
such as names, dates, teams, passwords the target used before, or other known facts. Attack-mode 4
then works like a normal attack, allowing the available compute devices to focus on one target with
all resources.

This needed a new token in the ruleset format. Where a trained grammar names a letter run and the
file of words it draws from, a hint ruleset names "H1", a slot filled at run time from what the
operator supplied. The included ruleset is not trained separately: "tools/pcfg_hints.py" derives it
from "default-passwords", and the same script turns any ruleset you trained yourself into one.

Attack-mode 9 reuses the "hints" ruleset for a different workflow. It obtains a separate set of
hints from each account name, hash line or module, then pairs those candidates with individually
salted hashes. It is intended to process many accounts at once rather than concentrate the attack
on one hash.

Fast hashes use device amplification. Slow hashes and modes without a suitable kernel use the host
engine.

See "docs/hashcat-pcfg.md".

3. A device-accelerated table attack in attack-mode 5

The table attack returns under its original attack-mode number. It applies token replacements to a
wordlist and amplifies the resulting combinations on the device.

Table entries can replace one or more bytes with any number of bytes, including an empty
replacement. Sources use longest-first matching. Multiple tables can be combined on the command
line, and the included tables cover leetspeak, case changes, emoji, sports templates and keyboard
layouts. With "identity=0" a layout table transliterates a word instead of varying it, which turns
a wordlist in one layout into what those keys produce on another, for every hash mode and every
attack the feed serves.

The "single=1" setting emits the original word once and then one matching substitution per
candidate instead of combining substitutions across every position. It does not use longest-first
matching because one substitution cannot overlap another. The "maxperm" and "cap=N" settings limit
substitution cost and the replacements retained for each source.

See "docs/hashcat-table.md".

4. Mask-controlled word placement with attack-mode 12

Attack-mode 12 places wordlists inside masks. The required "?w" marker inserts the first wordlist,
and the optional "?q" marker inserts a second wordlist after it. For example, "?w?d?d" is equivalent
to attack-mode 6 "wordlist.txt ?d?d" and "?w?q" is equivalent to attack-mode 1. The new part is that
both literal characters and masks can appear on either side of a word.

The mask is followed by the wordlist for "?w" and, when present, the wordlist for "?q". Mask files
are supported, but every mask must contain "?w". Increment mode is not supported because shortening
a mask can remove the word marker.

Attack-modes 1, 6 and 7 remain available and are translated internally to attack-mode 12 layouts.
Their candidate order and keyspace are unchanged. Restore files and status output retain the mode
entered by the user.

See the "Attack-Mode 12 Markers" table in "--help".

5. An account-aware association pipeline in attack-mode 9

Attack-mode 9 can now derive candidates directly from usernames and other per-hash information. It
splits account names at separators, case changes and letter-to-digit boundaries. Passwd input also
contributes the real name and the final component of the home directory.

Modules can provide other hints through a new module hook in which a plugin developer can implement
hash-mode specific extraction code. For example, mode 22000 supplies the network name and both
MAC addresses from a WPA capture.

This pipeline runs three phases by default: account words, the first 1000 rules from
"rules/rockyou-30000.rule", and the "hints" PCFG ruleset. The grammar phase is unbounded, so use the
"phases" setting when the run must finish. The existing one-word-per-hash input form keeps its
previous behavior.

Association attacks require one salt per hash. Unsalted modes and inputs that reuse a salt are
rejected. The wordlist form now stops on unusable lines instead of shifting all later
candidate-to-hash pairs.

See "docs/hashcat-association.md".

6. Seekable wordlists and a unified cache

Seek databases replace the dictionary cache. Periodic file offsets let a device start near its
assigned range instead of reading and discarding every earlier line.

Hashcat can seek directly in multi-block xz files produced by "xz -T0", multi-frame zstd files
produced by "pzstd", and concatenated xz streams. Single-stream files still require decoding from
the beginning, and gzip has no boundaries that support direct seeking.

Seek databases are stored in "cache/feeds/wordlist". The new "--cache-path" option relocates all
rebuildable data, including compiled kernels and feed caches, so a cluster can share it.

7. Brain server redesigned for large sessions

The brain server now divides each candidate database into 1024 shards selected from the high bits
of the candidate hash. Each shard has its own lock, sorted array and small overflow area for
commits. Commits and lookups therefore scale with the active batch instead of the full database.
The on-disk format remains readable, although the new session identities described below prevent
old files from being selected automatically.

Shard storage occupies one address range reserved for the session instead of many allocations that
grow and copy independently. Large sessions therefore use less processing time and resident
memory. The server reserves virtual address space rather than physical memory and must not run
under a virtual-memory limit.

The new "--brain-feed" option adds candidates from stdin to a running session without stopping the
server. See "docs/hashcat-brain.md".

8. Rules in every attack-mode

Rules were previously limited to attack-modes 0 and 9. Attack-modes 1, 3, 6, 7 and 12 now also
accept "-r" and "-g". Hashcat generates their candidates on the host and sends them to the device
as base words for rule amplification, matching the pipeline used by feed-based attack-modes 4, 5
and 8.

The existing mask processor still handles parsing, custom charsets and Markov tables. Mask files,
"--increment", "--markov-threshold" and custom charset options therefore behave as they do in
attack-mode 3. The "-j" and "-k" options still apply to the base and amplifier words rather than
the mask, and "--debug-mode" reports the rule that produced each candidate.

Host-side candidate generation costs fast hash modes most of their performance but has little
effect on slow modes, so hashcat reports the switch at startup. Larger rulesets amortize that cost
by doing more work per base candidate. Attacks without rules are unchanged and still generate
candidates on the device.

##
## Migration and Compatibility
##

These changes can alter scripts, distributed scheduling, file locations, monitoring output or
third-party plugins. If you maintain a module, bridge or feed outside the tree, or anything that
drives hashcat from a script, read this section before upgrading. Everything here is a change you
have to act on rather than one you will simply notice.

## Distributed work and progress

- Attack-modes 0, 6 and 7 treat multiple wordlists, directories, masks and increment lengths as one
  virtual keyspace; the left input to attack-mode 1 may now be a directory. The "--skip", "--limit"
  and "--keyspace" options address the complete run, so a distribution overlay can divide it
- Keyspace units still count base words before amplification. Rules and remaining mask positions
  multiply them afterward
- Separate dictionary rounds cannot be divided, so "--skip", "--limit" and "--keyspace" are rejected
  when several are queued, including multi-phase association. The "--limit" option is rejected with
  "--loopback", and both range options when "--restore" spans several masks
- Progress reporting is relative to the requested range in human-readable, JSON and
  machine-readable output. The "CURKU" and "Restore.Point" fields remain absolute positions in the
  complete keyspace, so monitoring integrations must no longer add "--skip" to "PROGRESS"

## Restore files

- The "--restore" option no longer executes the command line stored in a restore file, which a
  modified file could use to execute code. This is tracked as CVE-2026-68766. Hashcat prints a
  shell-quoted command to run instead, which uses "--restore-position" to keep the saved position
- Commands that cannot be represented safely for the current shell are shown as an escaped argument
  list. This includes non-printable bytes and Windows arguments containing "%", "!" or unsuitable
  quotes

## Runtime libraries and compressed input

- Compression support loads zlib, liblzma and libzstd on demand. Hashcat starts without them and
  reports a missing library only when it is required. See "docs/hashcat-compression-libraries.md"
- Attack-modes 1, 3, 6, 7 and 12, and therefore "--benchmark", require liblzma for the compressed
  "hashcat.hcstat2" table. Mode 11600 requires liblzma for LZMA or LZMA2 archives and zlib for
  DEFLATE, and without the library the attack runs but verifies nothing
- Wordlists, hash lists and rule files can no longer be read from ".zip". Gzip, xz and zstd remain
  supported
- Encoding conversion loads iconv only when requested. GNU libc and musl provide it directly, macOS
  and the BSDs provide system implementations, and Windows now uses native conversion APIs. Windows
  no longer supports EBCDIC, ISO-2022, "//TRANSLIT" or "//IGNORE", and rejects unsupported names

## Commands, devices, paths and tuning

- The "-c" and "--segment-size" options were removed with the old segmented wordlist reader. They
  now produce an unknown-option error, and "-c" will not be reassigned
- The "--dynamic-x" option takes the hash mode from the first line's "$dynamic_N$" tag. A missing or
  unknown tag is rejected, and later lines with another number are skipped with a warning. An
  explicit "-m" still wins and warns when it disagrees with the tag
- Included keyboard layouts moved from "layouts/*.hckmap" to "tables/layouts/*.table", so
  "--keyboard-layout-mapping layouts/de.hckmap" becomes "tables/layouts/de.table". The old files
  were removed, and a custom file in the previous format still works
- Installed builds no longer use "$HOME/.hashcat". Potfiles and sessions use the XDG profile
  directory, and rebuilt data uses the XDG cache directory. Existing files are not migrated
  automatically, but hashcat reports the old directory and both new locations
- All rebuilt data now sits below one "cache" directory: kernels under "cache/kernels" and feed data
  under "cache/feeds/<feed>". Scripts that clear compiled kernels must update their path
- The "--backend-devices-keepfree" option was removed. Hashcat uses free memory reported by CUDA,
  HIP, AMD OpenCL or hardware monitoring, and probes allocations if needed. The setting
  "HASHCAT_NO_AMD_FREE_MEM=1" ignores the AMD query; the desktop reserve is the last fallback
- Integrated GPUs are disabled by default when a discrete GPU is present. The "-d" option can
  select them explicitly, and systems with no discrete GPU are unaffected
- A device explicitly selected with "-d" now stops the run if it cannot start; "--force" continues
  with the remaining devices. Without "-d", failures are named and compatible devices continue
- Minimum backend and operating-system requirements changed; see "docs/hashcat-requirements.md".
  OpenCL 1.0 and 1.1 devices and the accelerator device type are no longer supported, and "-D 3"
  reports their removal instead of selecting accelerators
- The "-w" and "--workload-profile" options remain accepted but are ignored and omitted from
  "--help". Every run uses the former profile 3 launch budget, so a run that used "-w 1" or "-w 2"
  may gain throughput and lose desktop responsiveness
- The second ".hctune" column now identifies attack kernel 0, 1, 3 or 4 instead of attack mode. One
  row covers attack-modes 1, 6, 7 and 12. Custom mode 9 rows must use kernel 0, because value 9 is
  rejected. Included files label the column "Kern"

## Brain protocol and stored sessions

- The "--brain-client-features" defaults changed from 2 to 3, enabling both range and candidate
  deduplication. Hashcat automatically disables candidate storage for in-kernel modes with fewer
  than 1024 salts, while retaining range reservation. An explicit value overrides the decision
- The brain protocol changed from version 1 to 2. Candidate identities use a new hash and
  authentication uses stretched SHA-256. Clients and servers must be upgraded together, and a
  mismatch reports both versions during the handshake
- Existing brain databases are not converted. Their filenames no longer match the new session IDs,
  so they remain on disk without being loaded. An explicitly reused "--brain-session" starts with
  empty deduplication data

## Builds and third-party plugins

- The "SHARED=1" build setting is now the default on Linux and macOS. Android, the BSDs, CYGWIN and
  MSYS2 retain "SHARED=0". The Windows release is shared, but a native Windows build must request
  "SHARED=1" explicitly. The "SHARED=0" setting restores self-contained plugins
- Modules, bridges and feeds now link to one core library located above their plugin directory:
  "libhashcat.so.7", "libhashcat.7.dylib" or "hashcat.dll". Packages must ship it beside the
  frontend, and every third-party plugin must be rebuilt for interface 720
- The core now exports an explicit plugin API instead of every symbol used by an in-tree plugin.
  Host hash and cipher entry points are available even when no bundled plugin calls them
- Module porting removes "module_dictstat_disable", adds default assignments for
  "module_hash_hints", "module_length_sort", "module_usage_notice" and "module_advice_notice", and
  adds "hashcat_ctx" to the extra-parameter initialization and termination hooks
- The "hc_device_param_t" structure now stores kernels, programs and backend buffers in arrays
  indexed by "hc_dev_kern_t", "hc_dev_program_t" and "hc_dev_buf_t". Code using the old numbered
  members must be updated
- Bridge entry points now take "hashcat_ctx". Bridges must provide work-item granularity and the 7
  hardware-monitor hooks, or assign "BRIDGE_DEFAULT" when sensor data is unavailable, and the
  removed matching and forced-work-item flags have no replacement
- Bridges must use "bridge_salt_pos()" and "salt_per_pw" to select the per-candidate salt in
  association attacks
- Modules may declare a fifth auxiliary kernel with "OPTS_TYPE_AUX5". It reuses the bit formerly
  assigned to the removed "OPTS_TYPE_PT_ADD02" flag
- Native builds now target the build processor. Set "MAINTAINER_MODE=1" for distributable builds or
  "MCPU=<core>" to choose one explicitly. Raspberry Pi targets may be read from "/proc/cpuinfo",
  and mixed native and cross-compilation targets are rejected
- Rust plugins now require Cargo 1.85. Plugin 72000 requires Python 3.13 headers, and the bundled
  Python runtime was updated to 3.14.6
- Source builds now default to "PRODUCTION=0". Release packaging must set "PRODUCTION=1" to pin the
  release version; "SOURCE_DATE_EPOCH" makes repeated builds of that commit reproducible
- Build scripts must drop "USE_SYSTEM_ZLIB", "USE_SYSTEM_LZMA", "ENABLE_UNRAR",
  "USE_SYSTEM_UNRAR" and "USE_SYSTEM_XXHASH"; these settings are no longer recognized
- The plugin development guide at "docs/hashcat-plugin-development-guide.md" contains the complete
  core-library contract and the source changes required to port a plugin from v7.1.2

##
## New Hash Modes
##

- Added mode 36100: yescrypt
- Added mode 36200: gost-yescrypt
- Added mode 36300: scryptcrypt, scrypt (Unix)
- Added mode 35600: gost12512crypt [$gost12512hash$] (Unix)
- Added mode 34301: KeePass AESKDF (KDBX v4)
- Added mode 17050: GPG (AES-OCB-128 (SHA-1($pass)))
- Added mode 26150: Mozilla "key4.db" SHA384
- Added mode 35300: Kerberos 5, etype 23, TGS-REP (NT)
- Added mode 35400: Kerberos 5, etype 23, AS-REP (NT)
- Added mode 36800: OpenSSH Private Keys (bcrypt-pbkdf)
- Added mode 36400: KDE KWallet 4.13+ (PBKDF2-HMAC-SHA512, Blowfish)
- Added mode 36410: KDE KWallet < 4.13 (SHA-1, Blowfish)
- Added mode 36500: NTLM(NTLM_bin($pass))
- Added mode 35500: WordPress bcrypt(hmac-sha384($pass))
- Added mode 35700: phpass(md5($pass))
- Added mode 35800: Symfony Legacy SHA256

##
## New Features and Options
##

## Hash formats and recovered data

- Mode 23800 now uses hashcat's own RAR3 decoder. It supports both coding paths and all 6 standard
  RAR3 filters, and reports how many candidates reached an unsupported construct
- Mode 11000 accepts PrestaShop salts from 56 to 64 bytes. Its optimized kernels were removed
  because they cannot represent the longer salts
- Iteration counts allow 8 digits, up from 6, in modes 10000, 10900, 12000, 12100, 20200, 20300
  and 20400
- GPG (CAST5 (SHA-1($pass))) now supports nonzero IVs and "s2ktype" 3
- Added the dynamic hash plugin for mode 74000. A description in "--bridge-parameter2" can compose
  12 digests from MD2 through SHA-3, HMAC and PBKDF2-HMAC families, bcrypt, slicing, UTF-16LE, case
  conversion, and hex or base64 conversion
- Wordlists, hash lists and rule files now support zstd compression alongside gzip and xz.
  Multi-frame files produced by "pzstd" can be sought directly
- John's dynamic format now supports 56 built-in formats. The "--show", "--left" and "--remove"
  options preserve John's spelling
- Added "--hash-copy" to preserve each cracked hash exactly as supplied in the input when writing
  the outfile, potfile and status display
- Added "--encrypt-with-pubkey" to protect recovered plaintexts with an RSA public key of at least
  3344 bits. It loads OpenSSL 3 only when used, hides plaintext-bearing status fields, and refuses
  options that would write plaintext or restore state. See "docs/hashcat-encrypted-plains.md"
- The Windows build now opts into long paths and carries native version information

## Candidate generation and inspection

- Rule operands for "$", "^", "i", "v" and "o" can now contain non-ASCII characters directly
  instead of requiring one "\xHH" sequence per byte
- Added "--rules-concat" to concatenate several rule files instead of stacking them. The rule count
  is their sum rather than their product
- Added "--lookup" to report the position at which an attack reaches a candidate. It accepts
  "$HEX[...]" input and reports progress and an ETA for longer wordlist scans
- Rules now work in every attack-mode. Attack-modes 1, 3, 6, 7 and 12 generate candidates on the
  host and amplify them with rules on the device
- The "--debug-mode" option now reports the rule that produced a candidate in attack-modes 1, 3,
  6, 7 and 12, as it already does in attack-mode 0
- Stdin supports "--skip", "--limit" and "--restore" when the same stream is supplied again
- Attack-modes 1, 6 and 7 now support "--slow-candidates"
- Feed attacks support "--debug-mode" without "-r" or "-g" and describe their contribution in
  "Finding-Rule". The new "--debug-mode 6" writes the attack-mode 4 layout
- Added rule operations "h" and "H" for lowercase and uppercase hexadecimal conversion, "vNX" to
  insert "X" after every N characters, "S" to shift character case, and "BNX" to add byte X at
  position N

## Runtime control and reporting

- Added interactive rewind and advance controls on "r", "a", the arrow keys, "<" and ">". Holding a
  key increases the movement speed
- The "e" key can now extend or shorten a "--runtime" limit while the attack is running
- Added "--benchmark-pure" ("-B") for pure kernels. The existing "-b" benchmark continues to use
  optimized kernels, and every result names the kernel measured
- Added "--task-time-breakdown" for elapsed time by stage and peak memory usage. See
  "docs/hashcat-task-time-breakdown.md"
- Added "--pipeline-stats" for per-device feeding, copy, initialization, transfer, loop and
  comparison timings, including effective hash rate and machine-readable JSON output
- The "--cache-path" option relocates the complete cache and can point several hosts at shared data.
  Feed caches tolerate read-only storage and fall back to memory, but kernel-cache writes must
  succeed

## Extraction tools

- Added "tools/rar32hashcat.py" for mode 23800 RAR3 archives
- Added "tools/gpg-ocb-aes2hashcat.py" for mode 17050 GnuPG keys in "private-keys-v1.d"
- Updated "tools/mozilla2hashcat.py" for mode 26150 Firefox 146 "key4.db" files

##
## Performance
##

This section collects changes with a material effect on throughput, memory use or startup time.
Measurements depend on the device, runtime and workload.

- Brain sharding cuts commit, lookup and insertion time by 92 to 99 percent and resident-memory
  overhead by 95 percent
- PKZIP mode 17200 delivers 12.8 times the throughput on an RTX 4090 and gains 87 percent on an
  RX 7900 XTX
- DPAPI mode 15910 gains 19 percent on NVIDIA and 97 percent on AMD
- Kernels using at least 8192 bytes of private memory per item now run one warp or wavefront, making
  PKZIP 2.5 times faster on an RTX 4090 and 3.2 times on an RX 7900 XTX
- Fitting launch size from a few measurements, and reusing the lead device's result across
  equivalent devices and association rounds, cuts measured autotune time by 32 percent
- Rate-based launch selection prevents fast device-amplified attacks from choosing a starved
  one-unit launch inside the autotune tie margin, restoring 4.5 to 5.6 times the PCFG throughput
  on tested NVIDIA and AMD GPUs
- Candidate preparation now overlaps device execution in a separate thread. Brain clients retain
  their serialized reserve, lookup and commit path, and "HASHCAT_PIPE_SYNC" can select that path for
  every attack
- Candidate staging now clears only the first unwritten index rather than whole buffers, avoiding
  hundreds of megabytes of redundant memory clearing in large batches
- An unused event backlog was removed, saving 160 MiB of resident memory per attack
- The "--show" and "--left" options with "-o" no longer build, sort and copy per-hash records that
  their output handlers discard
- Devices that need the same kernel now coordinate through its cache filename and compile it only
  once
- Excluded CUDA devices no longer create a context merely to query memory, reducing backend setup
  time in heavily virtualized configurations
- Omitting an integrated GPU cuts cold mode 0 startup 26 percent at a 1 percent throughput cost
- Omitting a helper an OpenCL kernel never calls cuts cold builds 42 percent and backend-info
  startup 70 percent under ROCm 7.0.1
- Register-spill reserves sized from resident work items rather than the whole launch cut the
  reserve by 99 percent on a GeForce RTX 4090 running mode 0
- A unified-memory device now reserves the host buffers the run needs rather than half its pool,
  avoiding a fixed reserve that could leave nearly half the device unused
- Argon2 buffers are sized from the attack rather than the smaller self-test launch, allowing larger
  acceleration values without reserving unused memory
- Bitmap filters sized from the hash count and the device memory hierarchy, with probes mixing all 4
  digest words, make very large NTLM lists much faster and also help smaller ones
- Counting, parsing, sorting, deduplicating and indexing a hash list across all host cores makes
  large salted and unsalted lists much faster to load
- Duplicate hashes sharing one stored plaintext keep a duplicate-heavy list from costing excessive
  time and memory
- A hash list already completed by another run skips bitmap construction, kernel builds, device
  buffers and autotune. The "--outfile-check-dir" option is scanned before attack initialization
- Outfiles and potfiles are opened, locked and flushed once per launch rather than once per crack,
  and unsalted potfile lookups use digest-prefix buckets instead of scanning the complete list
- Sorting LM split halves and using binary search makes pairing scale near-linearly rather than
  quadratically
- Rule-file buffers that grow geometrically, a single "-r" file going straight to the amplifier, and
  reuse of the parsed rule make large rule files much faster to load and smaller in memory
- AVX2 or NEON line scanning and line-oriented stream locking accelerate initial wordlist and hash
  list scans
- Piped input supplied in blocks rather than one candidate at a time removes a major bottleneck in
  multi-device attacks
- Wide bridge units receive the complete iteration range per call and choose their launch size from
  2 measurements, avoiding repeated fill and drain work and a full autotune search
- Uncompressed wordlist seek-index scanning is more than 20 times faster, cutting measured database
  build time by 72 percent
- The "paw64" identity hash is 1.6 times faster than the xxHash it replaces in the brain and seek
  databases
- Parallel parsing, compact storage and cached unit tables cut large PCFG startup time by up to 95
  percent and memory use by more than 75 percent
- Hash-indexed table loading and linear-time longest-match indexing make construction of large
  generated tables scale linearly rather than quadratically
- The "--stdout" option applies rules across host cores with byte-for-byte identical output, avoids
  a device readback and uses a larger output buffer, making large rule workloads much faster
- Ordering candidates by length improves the 19 modes that request it by 50 to 200 percent.
  Modules opt in with "module_length_sort()", and "--length-sort-disable" turns it off
- Vector width 2 on Apple M-series GPUs improves 7 fast modes by 3.0 to 35.6 percent under Metal and
  OpenCL
- Reading "u32" words in the UTF-8 to UTF-16 decoder, rather than casting to bytes, improves 41
  modes by 2.3 to 9.0 percent
- One table lookup per AES round-key byte instead of 2 improves modes 16600, 22500, 22931, 22941,
  22951 and 23001 through 23003 by 15 to 41 percent on an RTX 4090
- Native AMD OpenCL byte operations improve Argon2 modes by 2 to 79 percent and mode 1700 by 27.6
  percent on an RX 7900 XTX
- Keeping yescrypt state in registers improves modes 36100 and 36200 by 13 to 45 percent, while
  local S-box placement can add another 39 percent
- Big-endian alignment using the same native operations improves mode 100 by 4.1 percent on an RX
  9070 XT with Mesa rusticl
- Compiler-generated byte-swap and bit-field code, in place of handwritten assembly, improves HIP
  modes 0, 1700 and 10800 by 0.4 to 0.7 percent
- A SHA-1-order shared hex table improves mode 18500 by 0.7 percent on an RX 7900 XTX
- Width conversion without the union type punning that Mesa rusticl lowered to scratch memory
  improves mode 6100
- Removing unnecessary 128-register limits lets modes 01800, 08300, 10700 and 16600 use the
  compiler's normal register allocation
- The native RAR3 decoder lowers per-device memory use compared with the former UnRAR path
- Straight wordlist attacks no longer reserve device memory for mask buffers they do not create
- Ending the outfile check thread without waiting for its polling interval reduces short-run
  shutdown latency
- Adaptive scrypt TMTO values beyond 2 improve underfilled, high-memory GPU configurations by 5 to
  46 percent without changing configurations that already fill the device
- Precomputing message templates for the md5crypt loop improves the optimized mode 500 kernel by 5
  to 33 percent across AMD, NVIDIA and Intel GPUs
- Removing always-zero message words from the optimized mode 0 kernels improves combinator
  attacks by 10 to 92 percent across AMD, NVIDIA and Intel GPUs
- Reporting the true RACF KDFAES iteration count instead of fixing "kernel_loops" at 8 lets
  autotune select the launch size, improving throughput by 27 percent on an RX 7900 XTX and 14
  percent on an RTX 4070 Ti Super
- Specializing the shared RC4 cipher for each caller improves 5 modes on an RTX 4090, including
  mode 9700 by 34.7 percent, mode 9800 by 30.4 percent and mode 10500 by 16.4 percent

##
## Improvements
##

## Backends: OpenCL, tuning and memory

- OpenCL clones share contexts and programs by default, divide the physical device's memory budget,
  and clean up together after a failed clone. Set "HASHCAT_CLSHARE=0" to restore separate contexts
- Runtime discovery now selects the newest versioned NVRTC, HIPRTC and HIP library, so Linux needs
  only runtime packages. It no longer probes nonexistent NVRTC names or derives HIP DLL names from
  the ROCm release in "HIP_PATH"
- HIP chooses its ABI from the loaded runtime, allowing one binary to support both ROCm and Hygon
  DTK
- Hardware monitoring now supports AMD GPUs under Mesa rusticl by obtaining their PCI address
  through "cl_khr_pci_bus_info". Temperature, fan and clock readings are now available
- Bridged modes accept "-n" and "--kernel-accel" values through 16384 candidates per launch. Other
  modes retain the 1024 limit
- Hardware and tuning diagnostics were expanded: NVIDIA throttling uses documented NVML reasons
  instead of NVAPI, out-of-range "-u" values name the accepted range, "HASHCAT_MEMORY" explains
  every device memory budget, and "-I" reports device cache size and bridge units
- Hashcat no longer sets "AMD_DIRECT_DISPATCH=0" for HIP, previously used to reduce host CPU load.
  On ROCm 7 it affects neither waiting thread CPU usage nor cracking speed, measured across several
  hash modes on a Radeon RX 7900 XTX
- The obsolete AMD setting "OCL_CODE_CACHE_ENABLE" and the old Intel CPU workaround
  "CL_CONFIG_USE_VECTORIZER" were removed. The "CUDA_CACHE_DISABLE" and "POCL_KERNEL_CACHE"
  settings remain for debug builds

## Autotune and tuning database

- The tuning database gained measured "kernel_accel" rows and NVIDIA GB10 tunings, and now matches
  vendor rows by backend vendor ID instead of individual device names
- Explicit module thread counts are preserved and reported. Mode 23800 now selects its count from
  the device, and Metal uses the compiled pipeline limit instead of capping every kernel at its
  32-thread SIMD width
- The "HASHCAT_AUTOTUNE2_VERBOSE" environment variable prints probe timings, the fitted cost model
  and the selected launch. See "docs/hashcat-tuningdb.md" for row selection and custom tunings

## Kernels

- The "HASHCAT_FORCE_NO_INLINE" environment variable builds kernels with "FORCE_NO_INLINE" for
  runtimes that compile inlined helpers slowly

## Attack modes and candidate input

- Wordlist keyspace counts physical lines. Progress includes unusable lines and reports them as
  rejected instead of shifting every later position
- The "HASHCAT_A0_LEGACY_READER=1" environment setting restores the previous attack-mode 0
  wordlist reader. Loopback and induction select it automatically because their inputs change by
  round
- Pure-kernel attack-mode 7 retains its mask-first path for one wordlist, while directories
  and multiple wordlists use the general path
- The "HASHCAT_SALT_INNER=1" environment setting tests each amplifier item against every active
  salt before moving to the next item, prioritizing the leading candidates when a run is
  interrupted
- Feeds can locate data in the shared directory and return "GENERIC_RC_SKIP" for a keyspace position
  that produces no candidate. The included "rust_random" feed is built when Rust is available

## Hash lists and potfile

- The "--show" and "--left" options support "--username" together with "--dynamic-x"

## Extraction tools

- Extraction scripts and modules now identify the corresponding hash mode or extraction tool more
  consistently

## Wordlists and compression

- Invalid "--encoding-from" or "--encoding-to" values now produce an explicit diagnostic instead of
  ending the run without explanation

## Brain

- Status, machine-readable and JSON output now split "Brain.Rejects" into position ranges reserved
  by another client and individual candidates already seen by the server

## Startup and diagnostics

- Kernel cache keys now use all 64 bits of "paw64". Existing entries that used the shorter key
  rebuild once
- Hashes supplied on the command line are now named in parse errors. Separator and token-length
  errors quote the failing text and state the expected separator or length
- Initial wordlist scans now report throughput in MiB/s

## Interface and status

- Pause loops poll more frequently, so quit and resume commands respond promptly
- Tab completion covers attack-modes 4, 5, 8 and 12, "--lookup", "--cache-path" and paths beginning
  with "$HOME"
- Status output now identifies wordlists as feeds, shows their position within a unified queue, and
  reports "Host Generator + PCIe" whenever candidates are not amplified on the device
- The "Restore.Sub" status field combines amplifier and iteration positions for 3 devices per
  update. Salt, amplifier and iteration counts moved into "Restore.Point" and appear only when
  greater than one
- Candidate windows, bridge speed lines and grouped temperatures now summarize large device sets
  instead of printing one row per device or unit on every update

## Benchmark

- Default "-b" modes now include yescrypt, gost-yescrypt, scryptcrypt, KeePass KDBX4 and Kerberos 5
  etype 23 AS-REP, and MyBB replaces vBulletin
- Removed redundant "md5($pass.$salt)" benchmark modes. The "md5($salt.$pass)" forms remain because
  their performance differs from the plain hash

## Plugins and bridges

- Bridge hardware sensors, unit names and "--hwmon-temp-abort" now refer to the hardware performing
  the work
- Mode 70100 allocates one "128 * r * N" scratch area per candidate, sizes batches from free host
  memory, reports "2 * N * p" iterations, and launches scrypt-jane in "-u" chunks so large N values
  show progress
- Mode 73000 names the required Python build and installation method. Python bridges use hashcat's
  log, and Rust unit information names the loaded plugin, for example "Rust [dynamic_hash.so]"
- Plugins can now receive bridge parameters, describe units made of several components, and call the
  core's device allocation, compilation and launch functions
- OpenCL hash and HMAC helpers read only the words covered by their input length, so plugins no
  longer need block-sized zero padding. Existing padded callers remain compatible
- The file layer adds "hc_fopen_mem()" for code that needs to read an existing memory buffer through
  the normal file interface
- Rust bridges share "hashcat-sys" and build every crate under "Rust/bridges" into "bridges/subs".
  Python bridge examples can save and reload the hashcat context for standalone debugging. CI now
  checks and tests both plugin families

## Modules

- Modes 8500, 14200, 16000, 22001, 27000 and 27100 now declare support for ASCII passwords only,
  which "--hash-info" identifies, other plaintext being shown as "$HEX[...]"
- Mode "-m 34300" is now named "KeePass Argon2 (KDBX v4)", which tells it apart from "-m 34301",
  the KDBX v4 AESKDF mode
- Usage notices now guide users of 45 modes, and mode 1000 can show an additional advice notice

## Package maintainers

- The "ENABLE_LTO=1" build setting enables link-time optimization and is used for releases. Linking
  the frontend and all plugins to one shared core reduces the unpacked release size by about 78
  percent
- Development installs contain the public plugin headers and an unversioned "libhashcat.so" for
  "-lhashcat". The "make install" target also installs feeds, rule files and PCFG rulesets in the
  shared data directory
- Added Android and Termux support, little-endian LoongArch, PowerPC and MIPS targets, and ARM
  support for the scrypt-jane bridge through sse2neon. OpenBSD now builds with its base compiler
- Apple Silicon builds are native by default. The "MACOS_UNIVERSAL_BINARY=1" build setting produces
  arm64 and x86_64 output on an Apple Silicon host and disables LTO
- Added the "hashcat_binaries" cross-compilation container and Ubuntu 26.04 CUDA and AMD runtime
  images. Image patches are isolated under "docker/patches", fail the build when they do not apply,
  and packaging images honor "HASHCAT_REF"
- CI builds FreeBSD and Pi 4 Linux ARM64 on every push, and OpenBSD, NetBSD and DragonFlyBSD weekly.
  Every artifact runs through "tools/test_package.sh" before upload
- Dependency tracking now rebuilds consumers of shared module bodies, feed headers and bundled
  libraries
- Bundled zlib and LZMA sources were removed in favor of system libraries
- Hashcat's RAR3 decoder replaces "deps/unrar" and removes the C++ compiler and runtime requirement
- The "paw64" hash replaces the bundled xxHash library

## Testing framework

- Added test oracles for 29 modes, including 4 GPG, 5 PKZIP and 6 collider modes. The "test.sh -S"
  command also cracks each mode's self-test vector when no independent oracle exists
- The "test.sh -g" command creates and cracks real containers for 105 LUKS, TrueCrypt, VeraCrypt,
  GPG, PKZIP, RAR, 7-Zip, PDF and OpenSSH modes while still running their normal oracles
- The "tools/test.sh" and "tools/test_edge.sh" scripts now cover attack-modes 4, 8, 9 and 12. The
  "-M" option selects 24 modes spanning every distinct path with vector width 1, and
  "tools/test_coverage.sh" requires every untestable mode to carry an explicit exemption
- Test modules may be Python or Perl, and generated passwords now include multibyte characters
- Added "tools/README.md", "tools/install_dependencies.sh" and "tools/requirements.txt" for test
  suite prerequisites. Perl module installation now reports each package that failed instead of
  only a count

## Documentation and licensing

- Added "docs/hashcat-rust-plugin-quickstart.md" for building a Rust bridge or feed
- Added "docs/packages.md" with what a package has to ship, the runtime libraries it depends on and
  the build settings a maintainer needs
- Rewrote the Camellia kernel from RFC 3713 under the project license with identical results in its
  16 modes. Added license texts for bundled Argon2, scrypt-jane, yescrypt, PPMd and the Windows
  liblzma and libzstd libraries
- Added "SECURITY.md", "CONTRIBUTING.md" and "AGENTS.md", and rewrote "README.md"
- Corrected the "--keep-guessing" section of "docs/limits.txt", which said valid passwords are
  always reported. Two candidates that match one hash inside one kernel invocation lose one of
  them, and the section now says so and separates the fast and slow hash cases

##
## Bug Fixes
##

## Memory safety: hash parsing and encoding

- Fixed hash-mode parsers reading beyond input bounds on a short or malformed hash line, in modules
  07100, 07900, 08500, 08501, 11600, 19600, 19700, 28800, 28900, 29100, 32100 and 32200
- Fixed salt iteration count underflow in many modules when the hash specifies zero, which could
  make the device hang
- Fixed 16 modules calculating salt iteration counts with a shift of 32 or more, beyond the type's
  width, which gave the wrong count
- Fixed "salt_repeats" underflow in modules 08900, 15700, 27700, 28200 and 29800
- Fixed scrypt kernels hanging at launch when parallelism is zero
- Fixed a one byte stack buffer overflow in 39 modules, where a salt buffer sized for the longest
  "generic_salt_encode()" output left no room for the terminator, which was then written just past
  it
- Fixed base64 salts overflowing "salt_buf" in "generic_salt_decode()" with non optimized kernels
- Fixed modules reporting encoded hash lengths larger than their output buffers, so callers wrote a
  terminator past the end. The "hash_encode()" function now limits the length it returns, and the
  potfile and status display paths that bypass it enforce bounds themselves
- Fixed a "0x80" padding byte being written beyond a full buffer, in modules 00111, 01411, 01711,
  05300, 05400, 11400, 13100, 13600 and 18200. The pure kernels of modes 00111, 01411 and 01711
  now also accept their advertised 256-byte salts
- Fixed encoding buffer overflows seen with "--show" and "--left" at the maximum data length, in
  modules 13100, 16900, 18200, 19600, 19700, 32100 and 32200
- Fixed a heap buffer overflow when constructing outfile lines, where a long username, hash or
  plaintext was copied in with no bounds check
- Fixed "--username" allocating only up to a NUL byte while recording the whole field length, so
  "--show" and "--left" read past the allocation and put unrelated heap data in the output
- Fixed 32 bit overflow in digest buffer offsets addressing the wrong hash beyond 4 GB
- Fixed a decode error on a damaged gzip, xz or zip file being taken as a byte count, overrunning
  hash list line counting and the brain attack hash and custom charset reader buffers
- Fixed a one byte overflow in "fgetl()" when a line exactly filled the caller's buffer
- Fixed hccapx handling in modules 02500, 02501, 22000 and 22001: compressed files no longer size
  decompressed arrays, ESSID and EAPOL lengths are validated, and short potfile lines are not read
  as 64 characters
- Fixed modules 16800 and 16801 reading short potfile lines past their end before checking the
  64-character PMKID
- Fixed KeePass parsers accepting invalid content lengths and writing oversized encodings in
  modules 13400, 29700 and 34300
- Fixed module 34300 indexing backward from the end of KeePass 4 hash lines shorter than 70 bytes
  before parsing them
- Fixed unbounded Kerberos account and realm copies in etype 23 modules 13100 and 18200 and etype 17
  or 18 modules 19600, 19700, 19800, 19900, 28800, 28900, 32100 and 32200
- Fixed module 25400 reading beyond a line's stated length with "%s", copying 32 bytes of user
  password past the end of it, using a stack variable after its scope ended, and appending a stray
  byte to a full 32 character PDF user password, which reached the hash line and the potfile
- Fixed heap buffer overflows in the PKZIP modules 17200, 17210, 17220, 17225 and 17230, where hex
  data tokens were decoded with no bounds check
- Fixed a decoded token exceeding the field it decodes into, in modules 11600 and 70000, and a
  decoded plaintext token doing the same in the RC4 modules 33500, 33501 and 33502
- Fixed writes beyond the hash list when an LM hash file grew during reading, where one line adds 2
  entries after space was checked for only one
- Fixed a stack buffer overflow with "--show" in split hash modes, where both plaintext halves were
  copied from the potfile into a 20 byte buffer unchecked
- Fixed buffer overflows in modules 08900, 10200, 21500, 21501, 28200, 70100 and 70200, where a
  base64 token without padding decodes to more bytes than its destination field holds
- Fixed a long password and salt overflowing the 128 byte packing buffer in the SAP CODVN F/G
  modes 7800 and 7801
- Fixed module 10000 accepting salts larger than its buffer, and its salt encoder leaving no space
  for a terminator
- Fixed a stack buffer overflow in the module 28600 encoder, which sizes its salt and base64 buffers
  for 32 bytes while the parser accepts 64
- Fixed module 26900 reading beyond the esalt on an SNMPv3 engine ID of nothing but zero bytes,
  and again while encoding for "--show" and the potfile
- Fixed mode 32300 accepting salts longer than its optimized kernel can hold, writing beyond its
  256-byte fields when joining fixed strings to salts near the maximum length, and overflowing the
  heap in its hash parser
- Fixed a base64 target hash overflowing the heap digest buffer in module 34000, and a base64 salt
  overflowing the stack decode buffer in module 23400
- Fixed the GPG modes 17010, 17020 and 17030 printing uninitialized bytes with "--show" and "--left"
  and writing them to the outfile, from a missing terminator in the hex encoded data field
- Fixed bridge scratch handling: modes 70100 and 70200 validate "N", "r" and "p", yescrypt sizes
  its buffers without 32-bit overflow, and scrypt-jane, yescrypt and Argon2 check allocation failure
- Fixed reads beyond buffer bounds in secp256k1 "point_mul_xy()" when the scalar is zero
- Fixed TrueCrypt and VeraCrypt modes leaking their container header on both error returns from
  "module_hash_binary_parse()", losing 512 bytes per rejected container across 36 modes

## Memory safety: candidates, rules and masks

- Fixed "-g" reading one past the end of its operator and character tables
- Fixed a stack buffer overflow for long words, where rule "/Xyp" asks "mangle_dupeblock_prepend()"
  for more bytes than its local buffer holds
- Fixed the rule engine's contain rejection reading beyond the candidate, and the 4 operators that
  reject on the first or last character reading past the end of an empty one
- Fixed missing bounds checks for class selector lookahead in the rule compiler
- Fixed stacked rule files corrupting memory or miscalculating device memory when a combined rule
  exceeds 31 functions or the rule count exceeds 32 bits
- Fixed a one-byte debug plaintext overflow when a rule produces a maximum-length password
- Fixed 33 combinator kernels reading beyond their 256 byte candidate buffers, where 2 words of up
  to 256 bytes each were combined with no check on the total length. Such a pair is now skipped
- Fixed a mask or custom charset over 1023 characters overwriting the length used to index it
- Fixed masks over 255 characters, and masks grown by a long custom charset, overflowing the 256
  byte buffers "--increment" uses to reverse and truncate them, and a heap buffer overflow with
  "--increment" and mask files
- Fixed "-g" drawing an index past the end of its selected group in 17 selection paths, and an empty
  "--generate-rules-func-sel" value causing division by zero instead of being rejected

## Memory safety: services and interface

- Fixed brain client and server overflows from negative peer lengths, more than 64 saved dumps, and
  more than 64 entries in "--brain-session-whitelist"
- Fixed keyboard mappings expanding TrueCrypt or VeraCrypt passwords beyond fixed buffers and
  mapping files with more than 256 entries overflowing their table
- Fixed unbounded reads from device memory for status candidates that could cause a crash
- Fixed the status display's rule file list overflowing its 4 KB buffer. The list is now truncated,
  because "snprintf()" returns the length it would have written
- Fixed "--outfile-check" reading and writing one byte beyond its line buffer
- Fixed reads beyond an unterminated restore directory field containing all 256 bytes
- Fixed a ".hctune" line with more than 7 fields overflowing the token array, which a user tuning
  file could reach
- Fixed an "snprintf()" length calculation error in the 7-Zip hash encoder for "-m 11600". No hash
  the parser accepts is long enough to trigger it
- Fixed substantial allocations retained by early returns from "outer_loop()" and the "--identify"
  probe, including rejected option combinations
- Fixed "profile_dir", "session_dir" and "shared_dir" leaking when hashcat runs from its install
  directory
- Fixed the argument vector loaded from a restore file leaking once per libhashcat session

## Hash modes and kernels

- Fixed modes 22000 and 22001 refusing the old PMKID format when it was written with ":" separators
  rather than "*". The line was rewritten into the current format by copying it whole, which left all
  4 fields as one token
- Fixed mode 3000 silently truncating wordlist entries longer than 7 characters instead of rejecting
  them
- Fixed "-m 19800" and "-m 19900" ignoring the salt field, so a principal whose salt is not the
  realm followed by the user never cracked, and a hash line carrying that field would not load
- Fixed the 41 modes that convert UTF-8 to UTF-16 in the kernel rejecting a correct password holding
  the euro sign, Japanese kana or most other 3 byte characters, where only 4 of 16 valid 3 byte lead
  bytes and 3 of 5 valid 4 byte lead bytes were recognized
- Fixed "-a 3" failing to crack a password with characters outside ASCII, where each UTF-8 byte was
  widened into a separate character rather than decoded, in modules 00130, 00140, 00170, 01430,
  01440, 01470, 01730, 01740, 01770, 10830, 10840, 10870, 13500, 13800, 29000 and 29200
- The "-a 3" fix for characters outside ASCII needs "--backend-vector-width 1" on CPU devices
- Fixed a missing barrier in the module 22100 loop kernel, which gave wrong results once workgroup
  threads left lockstep, because a thread starting the next round could overwrite shared
  precomputed salt data while the others were still reading it
- Fixed "--keyboard-layout-mapping" having no effect in the 15 TrueCrypt and VeraCrypt container
  modes 6241-6243, 13721-13723, 13741-13743, 13761-13763 and 13781-13783. Binary container parsing
  now runs "module_hash_decode_postprocess()" like text hash parsing
- Fixed the VeraCrypt modes 13723, 13733, 13751, 13753, 13761 and 13763 failing their self test,
  and hanging under "--self-test-disable" on the password they had just found, by keeping the 15
  XTS header helpers out of line. TrueCrypt, DiskCryptor and LUKS were never affected
- Fixed module 01500 cracking nothing in "-a 3" with multiple hashes sharing one salt, where the
  path that compiles the salt into the kernel had an empty stub for multiple hashes
- Fixed stored DEFLATE blocks in PKZIP modes 17200, 17220 and 17225 on AMD HIP and OpenCL, and in
  the attack-mode 1, 3, 6, 7 and 12 kernels for mode 17230
- Fixed module 10500 rejecting PDF hashes with an empty ID field
- Fixed "-m 24420" rejecting the self-describing "$PEM$2" line that "pem2john.py" emits for any
  PRF other than SHA-1, so a key from the current script could not be loaded at all. A line
  naming another KDF or PRF is now refused rather than read as SHA-256
- Fixed GPG modes 17010, 17020, 17030, 17040 and 17050 failing at particular iteration counts and
  password lengths because stale salt bytes entered key derivation. Mode 17030 also reads AES-128
  keys from the SHA-256 state with the correct word width
- Fixed modes 7800 and 7801 truncating a SHA-1 message past byte 119, so a matching candidate did
  not crack
- Fixed the SNMPv3 modules 25000, 25100, 25200, 26700, 26800, 26900 and 27300 rejecting an engine
  ID shorter than 13 bytes, and 26900 and 27300 padding every ID to 17 bytes with zeros, so a
  shorter one did not crack during password localization
- Fixed the 8 scrypt modules reporting an incompatible self test, or a mixed configuration, for a
  file holding a single hash, where large "r" and "p" values made derived buffer sizes overlap error
  flag bits. Such hashes are now rejected for their memory requirements
- Fixed modules 03710, 04010 and 04110 advertising a salt limit past what the optimized kernel
  holds, so a 24 to 51 byte salt was accepted under "-O" and never cracked
- Fixed "-m 35100" rejecting a password longer than 15 characters, having no pure kernel and
  therefore always using the optimized one
- Fixed association workgroups reusing one hash's salt, keyboard mapping or page table, and host
  hooks receiving only the launch salt. The fixes cover modules 01100, 05300, 05400, 05600, 07300,
  08000, 11600, 13800, 17200-17230, 23800, 06211-06233, 13711-13773 and 22100
- Fixed the PKZIP modes 17220 and 17225 overrunning the private-memory aperture on AMD GPUs, which
  could prevent HIP launches or severely reduce throughput
- Fixed mode 31100 hanging at vector width 16 in the optimized "-a 3" kernel on AMD GPUs, where the
  SM3 message schedule grew large enough for the compiler to relax a branch through the register
  pair holding the return address. The function returned into itself, and is now inlined
- Fixed secp256k1 "inv_mod()" looping forever on zero input and hanging "clFinish()" on CPU devices
- Fixed the pure kernels of modes 01500, 03000, 08500, 14000, 14100 and 16000 using shared memory on
  a CPU device, which has none of its own
- Fixed modules 8500, 8501, 12400 and 14100 failing to compile on OpenCL runtimes that report local
  memory as global memory
- Fixed "-m 17220 -a 3" build failures from including "inc_checksum_crc32.cl", renamed during CRC
  restructuring
- Fixed "-m 73000" crashing after a run under Python 3.14, where its bridge released 4 Python
  references it did not own and leaked 2 tuples
- Fixed 11 hash modes writing "--separator" between internal hash fields, which hid where the
  plaintext starts when an outfile line is split on that character, and module 30700 writing fields
  with "--separator" when its parser expects a colon
- Fixed the hash category for Mega.nz links protected by passwords
- Fixed a module file that exists but cannot be loaded crashing autodetection, "--identify",
  "--help" and "--hash-info", and memory leaks in "module_load()" when the file cannot be opened or
  carries no "module_init()"

## Backends: OpenCL, tuning and memory

- Fixed the Windows HIP runtime version check accepting every version because it compared
  incompatible version scales
- Fixed OpenCL 1.2 devices requiring extension strings for atomics and byte-addressable stores, and
  Beignet or Mesa version strings triggering the obsolete unstable-driver rejection
- Fixed "-cl-std=" being chosen from the platform's OpenCL version rather than the device's, which
  also made the too-old diagnostic name the wrong version
- Fixed an optimized kernel built with rules never starting on Apple Silicon under OpenCL at
  "--backend-vector-width 2", where Apple's compiler inlined 64 of the 66 rule functions into one
  body it never finished. They are kept out of line there, as they already are on the Metal backend
- Fixed OpenCL work sizes exceeding what a device allows for the first work item dimension,
  which hashcat never asked about before
- Fixed devices of the same class sharing one acceleration, thread count and batch size, which could
  exceed what one device's buffers hold. Classification ignores memory, and a card driving a display
  allocates less, so shared values are now capped by each device's own maximums
- Fixed OpenCL 1.2 platforms compiling as OpenCL C 1.1 and rejecting static data at file scope
- Fixed a compiled kernel being readable before its writes completed, where the writer truncated the
  file before locking it and readers took no lock. Another hashcat process could load a truncated
  kernel, so kernels are now written to a temporary file and renamed into place
- Fixed a device being skipped when hashcat could not parse its OpenCL driver version, which was
  treated as too old. The version is now reported once and the device stays available
- Fixed a device being dropped when a tuning row pinned an acceleration value too large for its
  memory. The pinned value is now lowered, but only after every other fit has been tried
- Fixed autotune deriving an invalid acceleration for one-thread kernels and using misaligned
  thread-count steps across wavefront boundaries
- Fixed underflow in the scrypt and argon2 available memory calculations, where a device reporting
  less memory than the reserve asked for an enormous allocation
- Fixed OpenCL kernel lookup returning "aux1" for all 4 auxiliary kernels, so register spill
  reserves were calculated from "aux1" alone
- Fixed OpenCL launching "vector_width" times too many work items in the "loop2" kernel, in modes
  14800, 15310, 15910, 27500 and 27600
- Fixed per-work-item buffers being sized from the unrounded launch count, which let vectorized
  kernels read and write past their ends
- Fixed disabling a mode's JIT kernel cache also disabling the shared, Markov and amplifier caches
- Fixed hashcat continuing after a required thread failed to start, and joining an uninitialized
  thread handle afterwards. A worker that cannot start now runs its work on the calling thread, and
  a cracking, monitor or outfile check thread that fails to start is reported

## Backends: Metal and resource lifecycle

- Fixed a CUDA benchmark retaining register-spill allocations between modes
- Fixed 6 backend leaks: the NVRTC and HIPRTC option arrays, a string returned by
  "module_jit_build_options()", an autotune event and 2 Metal strings
- Fixed larger kernels, mostly the SHA-512 modes, failing to compile on Metal, which now builds
  its libraries optimized for size
- Fixed Metal leaking one placeholder buffer per launch and 25 pipeline objects per device. The
  "run_kernel()" function now releases the placeholder it creates for an unused argument, as
  "run_kernel_amp()" already did
- Fixed a failure to read the Metal runtime version ending the whole macOS run, and disabling the
  OpenCL devices with it. It now disables only Metal and falls back to OpenCL
- Fixed crashes on macOS Tahoe when maximizing concurrent Metal pipeline compilation
- Fixed the AMD "V_PERM_B32" probe indexing with the device total instead of the loop counter when
  it reuses a neighboring device's result, so it compared with the last enumerated device rather
  than the preceding one and could take a result from an unprobed slot, disabling support
- Fixed the HIP blocking sync flag reaching only the first of several AMD GPUs, because
  "hipSetDeviceFlags()" affects the currently selected device and the flag was set before selecting
  one. Hashcat now selects each device first
- Fixed backend sessions remaining open after startup failures outside benchmarks
- Fixed CUDA contexts remaining on the main thread after creation, session cleanup freeing buffers
  under the wrong device context, and context pops writing into device-thread handles

## Backends: bridges and hardware monitoring

- Fixed bridged modes autotuning an empty loop placeholder instead of measuring bridge work
- Fixed the first bridge unit's tuning result overwriting every other unit even when their widths
  and performance differed
- Fixed wide bridge units sizing launch buffers as though "kernel_accel" counted candidates instead
  of waves, which could underallocate them and cause an illegal memory access. The "-n" option now
  counts waves, and bridge speed lines report "Accel" instead of "Batch"
- Fixed the Python bridges dropping a borrowed reference and then returning a pointer into it, and
  early returns leaving the thread state attached, which ended the process on the next attach
- Fixed a bridged hash mode ignoring "-d" and running every bridge unit, and accepting a device
  number that does not exist without saying so
- Fixed a Metal device being disabled when "-d" also picked an OpenCL device, with no message
  saying how to select Metal instead
- Fixed runs exiting successfully after losing every device, most visibly during benchmarks
- Fixed "-u" values above 1024 being rejected even though 1024 is only the default limit
- Fixed a bridged hash mode accepting "-n" and then ignoring it, where the launch size was
  overwritten with the candidate count the bridge advertises, the check having overlooked the modes
  whose limit is well above 1024
- Fixed a backend initialization error on every run where CUDA is installed with no NVIDIA GPU, or
  HIP with no AMD GPU
- Fixed "-b --benchmark-all" crashing at the first hash mode requiring a bridge, where bridge hooks
  were called without a loaded bridge
- Fixed the "-Y 4" setting printing the same hardware-monitor data 4 times for one physical card
- Fixed hardware monitoring probing unsupported sensors on integrated Intel GPUs or devices without
  PCI addresses, testing the wrong vendor handle, and showing only the first bridge unit
- Fixed a HIP device property failure being reported as a "hipDeviceGetAttribute()" error, without
  naming the device it happened on
- Fixed outdated driver messages naming Intel regardless of device vendor, where the checks cover
  Intel CPU devices and AMD, NVIDIA and Apple GPU devices. The message now names the device's vendor
- Fixed "-I" listing one backend device per bridge unit rather than per physical device
- Fixed a modified kernel not being recompiled, because the cache key omitted the kernel source
- Fixed Apple IOKit initialization running twice and leaking an SMC connection and an IOReport
  subscription on every run

## Attack modes and candidates

- Fixed "--stdout" writing a candidate that contains a line ending as two lines. Candidates now use
  the same "$HEX[...]" form as the outfile, so one candidate is always one line
- Fixed "--generate-rules-seed" having no effect outside Windows, where the seed went to "srand()"
  while "-g" drew from "random()", so every run produced the same rules
- Fixed 3 attack-mode 9 paths launching more work items than salts, which could read beyond the salt
  buffer and abort on small hash lists
- Fixed bridged association attacks reusing one launch salt. Modes 70000, 72000, 73000 and 74000
  previously recovered only one hash, while 70100 and 70200 failed when per-hash costs differed
- Fixed 2 large rule files incorrectly reporting completion after a keyspace of 65536
- Fixed checkpoint cancellation after a device had already stopped, which could discard undispatched
  keyspace and delete the restore file. Restore positions beyond the keyspace no longer wrap to an
  enormous remaining count
- Fixed rejected words under "--slow-candidates" shifting later positions. Attack-mode 1 now applies
  case and encoding conversion before rejecting oversized amplifier words, and attack-modes 1 and 7
  no longer reuse stale candidates after a rejection
- Fixed a mask file line using all 8 custom charsets keeping the previous line's mask, where only
  the first 8 of the 9 fields had their lengths reset
- Fixed the Danish keyboard layout, which contained no valid mappings, and reject empty mapping
  files or files with multiple replacements for one source before the attack starts
- Fixed host title-case classes differing from GPU rules, memory rules padding candidates before
  "M", and "-g" omitting 9 generated rules by treating parameterless operators as positional
- Fixed silent write failures with "--stdout", where a full disk or closed pipe lost candidates with
  no error
- Fixed the remainder of an oversized stdin line being emitted as a candidate. A line exceeding the
  reader's buffer is now discarded, as wordlists already did, because no password is that long
- Fixed a file handle leak in pure kernel "-a 7", once per mask, dictionary and device, and rule
  file buffers leaking when a ruleset was loaded
- Fixed "--keyspace" and "--total-candidates" being accepted with attack-mode 9, where the hash
  file was opened as a wordlist
- Fixed "--stdout" printing candidates out of order when several devices each printed their slice,
  though the stream still held every candidate exactly once. A "--stdout" run now uses one device,
  while cracking runs still use all of them

## Hash lists, potfile and outfile

- Fixed combinator and hybrid attacks reporting a plaintext truncated to 256 characters, which did
  not match the cracked digest
- Fixed "-a 7" writing malformed outfile records for a candidate longer than 256 bytes, where the
  outfile cursor advanced beyond the hex output actually written
- Fixed "--show" corrupting LM output for a hash with an unusable right half, and incorrect split LM
  hash pairing
- Fixed "--outfile-check" racing with cracking threads on digest counters
- Fixed an undeletable induction dictionary being rediscovered and read indefinitely after use.
  Induction now stops and reports the reason
- Fixed scanned induction dictionary paths leaking on every rescan, the loopback filename leaking
  once per round, and the 4 KB loopback buffer allocated at startup never being freed, which
  together accumulated throughout a "--loopback" run
- Fixed "--slow-candidates" reporting a crack position relative to the batch rather than the
  keyspace, which "--outfile-format 4" then wrote as a position from the launch start

## Brain

- Fixed a 32 bit overflow in the brain session hash
- Fixed partial runs colliding with the full attack in the brain database because attack identities
  did not include "--skip"
- Fixed the brain server not saving databases on "SIGTERM" and not flushing redirected logs after
  each line. Service and container stops now preserve both session data and final diagnostics
- Fixed a brain client carrying on without deduplication when it could not reach the server. A
  failed connection now reports an error naming the host and port
- Fixed "--brain-session-whitelist" crashing when given a value that identifies no session
- Fixed attack-mode 0 clients launching short batches after another client's reservation, and
  "-S" clients under attack-modes 1 and 3 counting reserved keyspace twice
- Fixed the brain server reporting huge read or write sizes for an empty attack database

## Interface and status

- Fixed "--bitmap-max" accepting unusable values above 28 instead of rejecting them
- Fixed the startup "Bitmaps" row reporting a rotation count for shifts removed from the kernel
  parameters
- Fixed "--runtime" having no effect with "--stdout", because the monitor thread that enforces it
  only started when "--stdout" was off
- Fixed the final status struct being overwritten once per round rather than released, and a
  "hashinfo_t" allocated twice and freed through a field already cleared
- Fixed "POSIXLY_CORRECT" changing which arguments hashcat reads as options, silently handing its
  own options to the attack
- Fixed diagnostics being written into the candidate stream under "--stdout", where they were read
  back as candidates
- Fixed "--speed-only" measuring with the bitmap reduced to one entry, which reported incorrect
  speeds for a large hash list
- Fixed "--progress-only" measuring with the bitmap filter forced to one bit, which substantially
  understated the progress step for large hash lists and ignored "--bitmap-min" and "--bitmap-max"
- Fixed "--progress-only" reporting speed hundreds of times too low when a fast attack required
  several base-word batches, because it paired the final batch count with all elapsed time
- Fixed messages from different threads overwriting one another, where every message was formatted
  into one shared buffer. That buffer now has its own lock, because a log event cannot take the
  event mutex handlers already hold
- Fixed "No devices found/left." saying neither why the devices were removed, nor that bridge units
  existed with no device to drive them
- Fixed "--backend-devices-virthost" values past the last device reporting no devices found. The
  error now identifies the nonexistent device number
- Fixed "--help" examples and "docs/hashcat-brain.md" naming "rules/best64.rule" instead of
  "rules/best66.rule", so a copied command failed

## Windows

- Fixed paths and arguments being converted through the system code page. Hashcat can run from any
  Unicode directory, generated PID, tuning and potfile paths remain valid, and masks or rules no
  longer arrive with unsupported characters replaced by "?"
- Fixed the working directory being searched for libraries loaded by name, so a DLL placed beside a
  wordlist can no longer be loaded. Libraries supplied beside "hashcat.exe" are unaffected
- Fixed mode 23800 importing "libstdc++-6.dll" and "libgcc_s_seh-1.dll", which prevented it loading
  in release and MSYS2 builds without those libraries on "PATH"
- Fixed file locking being an unimplemented stub, which let 2 hashcat processes interleave lines
  when appending to the same potfile
- Fixed Windows thread entry points returning the wrong width and leaking handles on joins and host
  hooks, including frequent leaks in mode 11600
- Fixed Python plugins disappearing when the bundled runtime was not 3.12, and native MSYS2 builds
  of bridges 70200 and 70300 missing the Windows Python include path
- Fixed missing Windows version information in module 74000's Rust bridge "generic_hash.dll", which
  caused antivirus false positives
- Fixed console messages using the console code page instead of UTF-8, so non-ASCII output no longer
  requires "chcp 65001"
- Fixed stderr message colors being applied to the console used by stdout, where terminal detection
  checked stdout regardless of the stream being written
- Fixed hashcat having to be started from its own folder, where the path of the running binary was
  never read and every folder resolved against the working directory. A Unicode folder still has to
  be the working directory on CUDA, whose runtime compiler cannot open such an include path
- Fixed a UNC install path leaving the kernel include path empty, where the leading "//" of
  "//server/share" inside a "-D" value was read as the start of a C comment

## Build and packaging

- Fixed archives built from development commits carrying the 7.1.2 release name. Package archive
  names now use the version stamped into the built binary
- Fixed "DEBUG=2" failing to link with AddressSanitizer
- Fixed a failed build step leaving partially written output behind, which a later "make" then
  treated as up to date
- Fixed "make linux" building every module, bridge and feed against the native core instead of the
  Linux one
- Fixed "make" failing to find Rust libraries when "CARGO_TARGET_DIR" is set in the environment
- Fixed argon2id and scrypt bridge build failures with "-march=native", which selected a processor
  the compiler then rejected, on DragonFly BSD and under some hypervisors
- Fixed "extra/tab_completion/install" finding the repository root one directory short, writing the
  literal "${ROOT_PARENT}" as the hashcat path, rerunning itself under "sudo" and appending the
  completion to root's ".bashrc" rather than yours
- Fixed the NetBSD build conflict between its "qsort_r()" declaration and "sort_r.h", and a "cpuset"
  warning there
- Fixed both Python bridge makefiles defining 2 recipes for one target, so "make" reported an
  overridden recipe on every build without the Python headers
- Fixed 4 redefinition warnings in the Python bridges, where "Python.h" was included after the
  headers whose feature macros it sets itself

## Tools and tests

- Fixed "luks2hashcat.py" writing an invalid hash file with keyslot names in its output, and
  treating LUKS2 reencrypt keyslots as password keyslots
- Fixed "tools/test.sh -p -I" reading the wrong hash type from a run directory, so the packaged
  script it produced ran no tests
- Fixed "tools/test_edge.sh" emptying the hash file for a hash containing a single quote, which
  failed every combination
- Fixed "tools/test.pl" quoting fields without escaping their contents, and "test_edge.sh" decoding
  those fields with "eval". Words, salts and hashes are now passed hex encoded
- Fixed Perl test modules for modes 30901 to 30905 calling deprecated "Bitcoin::Crypto" symbols
- Fixed the edge test suite not telling 43 test modules which kernel was under test, causing every
  module to generate optimized-kernel vectors even when the pure kernel was under test
- Fixed edge tests for binary-container modes 5200 and 9000 passing hashes on the command line, and
  a 1-character mode 99999 hash being treated as an empty vector
- Fixed multi-hash tests accepting swapped hash and plaintext pairs, and edge tests reading cracks
  from stdout where warnings could mix into the results
- Fixed test modules 28504 and 30904 generating no P2WPKH hashes from uncompressed keys, and mode
  30906 declaring a password limit below its own vectors
- Fixed the test dependency installer using a PyPI "pycryptoplus" package incompatible with current
  setuptools. It now verifies every Python dependency after installation
- Fixed test modules 14400 and 15000 deriving password lengths from the wrong budget, and the mode
  25700 oracle treating non-ASCII candidate bytes as signed values

* changes v7.1.1 -> v7.1.2

##
## New Algorithms
##

- Added hash-mode: Generic Hash [Bridged: Rust] with the new Rust Bridge
- Added hash-mode: Besder Authentication MD5

##
## Improvements
##

- CRC Libraries: CRC restructure
- GOST Libraries: GOST R 34.11-94 restructure
- SAP CODVN H: Raised the maximum salt length for hash-mode 35000 (isSHA512) from 16 to 192

##
## Bugs
##

- Argon2 Libraries: fixed call to vector helper functions when only using scalar buffers
- Argon2 Libraries: fixed esalt access pattern across Argon2, LUKS2, and KeePass KDBX4 kernels
- Argon2 Libraries: fixed false negatives for password and salt combinations longer than 64 characters
- Folders: fixed hc_mkdir_rec(), which was not creating directories correctly
- Status View: fixed broken backward compatibility in machine-readable mode introduced in v7.1.0
- sha224(sha224($pass)) and sha224(sha1($pass)): Fix hash category as unsalted

* changes v7.1.0 -> v7.1.1

##
## Improvements
##

Docker: Add initial support for AMD GPUs when running Hashcat inside Docker using HIP and OpenCL
Hardware Monitor: Suppress fan speed warnings for unified memory setups (typically internal GPUs)
Testing Framework: Enable hardware monitor to improve testing of memory-heavy algorithms
Unit tests: Updated install_modules.sh to use an external module for Digest::MD6

##
## Bugs
##

- Fixed multi-hash handling in LUKS2 and KeePass (KDBX4), while single-hash remained unaffected

* changes v7.0.0 -> v7.1.0

##
## New Algorithms
##

## Endpoint

- Added hash-mode: AS/400 DES
- Added hash-mode: AS/400 SSHA1
- Added hash-mode: Blockchain, My Wallet, Legacy Wallets
- Added hash-mode: Cisco-ISE Hashed Password (SHA256)
- Added hash-mode: Keepass4
- Added hash-mode: SAP CODVN H (PWDSALTEDHASH) isSHA512
- Added hash-mode: sm3crypt $sm3$, SM3 (Unix)

## Primitives, KDFs, Libraries

- Added hash-mode: BLAKE2b-256
- Added hash-mode: MD6 (256)

## Constructs

- Added hash-mode: BLAKE2b-256($pass.$salt)
- Added hash-mode: BLAKE2b-256($salt.$pass)
- Added hash-mode: sha224($pass.$salt)
- Added hash-mode: sha224($salt.$pass)
- Added hash-mode: sha224(sha1($pass))
- Added hash-mode: sha224(sha224($pass))

##
## Improvements
##

- Attack-Modes: Use 64-bit counters for amplifier keyspace
- Host Memory: Update method to query free host memory using /proc/meminfo when available
- Docker: Add initial support for running hashcat inside Docker
- Device Memory: Warn instead of waiting on high GPU memory usage
- Backends: Enhanced AMD Windows OpenCL and HIP compatibility for legacy models
- Rules: Add early syntax check for -j and -k rules
- Device Memory: Do not disable hwmon interface by default in speed-only and progress-only mode
- JSON output: Refactor Bridge unit reporting for clarity
- Backends: Ignore devices from Microsoft OpenCL D3D12 platform
- Building: Add MAINTAINER_MODE flag to disable hardcoded CPU optimization flags
- Bash: Add missing parameters to bash completion script
- Dependencies: Downgraded unrar source from 6.2.7 to 6.0.5
- Benchmark: Update default hash-mode selection
- Libraries: DES and Blowfish restructure
- Building: Add initial support for OpenBSD and DragonflyBSD

##
## Bugs
##

- Fix broken JSON formatting when using the --status-json flag
- Fix issue where -k single rules were ignored when used with -a 1 and -S

##
## Python Bridge
##

- Fix unsalted hashlist support
- Fix the esalt structure, it was too large
- Improve support from 1:1 password-to-hash to 1:N password-to-hashes
- Improve stand-alone debugging of Python Bridge stubs
- Improve salt/esalt debugging of Python Bridge stubs
- Enable potfile output by default for -m 73000 and 72000

##
## Other changes
##

- Argon2: Add argon2_init() wrappers to support private address space
- Change hash-category for hash-modes 25600, 25800, 28400 and 30600 to HASH_CATEGORY_RAW_HASH_SALTED
- Suppress optimization advice in final approach mode
- Backend: Split backend_session_begin() into smaller compute-runtime specific functions
- Sanity check: Abort if a custom-charset is defined on the command line and a mask file is used
- Building: Disable Argon2 optimized path on RISC-V
- Shared: improved 32-bit and 64-bit add and multiply overflow helper functions

* changes v6.2.6 -> v7.0.0

##
## New Algorithms
##

## Primitives, KDFs, Libraries

- Added hash-mode: AES-GCM
- Added hash-mode: Argon2 (Argon2i/d/id)
- Added hash-mode: BLAKE2s-256
- Added hash-mode: CAST
- Added hash-mode: PBKDF1-SHA1
- Added hash-mode: RC4-40/72/104/128
- Added hash-mode: RIPEMD-320
- Added hash-mode: ShangMi 3 (SM3)
- Added hash-mode: HMAC-BLAKE2S (key = $pass)
- Added hash-mode: HMAC-RIPEMD160 (key = $pass)
- Added hash-mode: HMAC-RIPEMD160 (key = $salt)
- Added hash-mode: HMAC-RIPEMD320 (key = $pass)
- Added hash-mode: HMAC-RIPEMD320 (key = $salt)

## Bridged

- Added hash-mode: argon2id [Bridged: reference implementation + tunings]
- Added hash-mode: scrypt [Bridged: Scrypt-Jane ROMix]
- Added hash-mode: scrypt [Bridged: Scrypt-Yescrypt]
- Added hash-mode: Generic Hash [Bridged: Python Interpreter free-threading]
- Added hash-mode: Generic Hash [Bridged: Python Interpreter with GIL]

## Endpoint

- Added hash-mode: 1Password, mobilekeychain (1Password 8)
- Added hash-mode: Adobe AEM (SSPR, SHA-256 with Salt)
- Added hash-mode: Adobe AEM (SSPR, SHA-512 with Salt)
- Added hash-mode: Anope IRC Services (enc_sha256)
- Added hash-mode: Apache Shiro 1 SHA-512
- Added hash-mode: BestCrypt v4 Volume Encryption
- Added hash-mode: Bisq .wallet (scrypt)
- Added hash-mode: Bitcoin raw private key (P2PKH), compressed
- Added hash-mode: Bitcoin raw private key (P2PKH), uncompressed
- Added hash-mode: Bitcoin raw private key (P2SH(P2WPKH)), compressed
- Added hash-mode: Bitcoin raw private key (P2SH(P2WPKH)), uncompressed
- Added hash-mode: Bitcoin raw private key (P2WPKH, Bech32), compressed
- Added hash-mode: Bitcoin raw private key (P2WPKH, Bech32), uncompressed
- Added hash-mode: Citrix NetScaler (PBKDF2-HMAC-SHA256)
- Added hash-mode: CubeCart (whirlpool($salt.$pass.$salt))
- Added hash-mode: DANE RFC7929/RFC8162 SHA2-256
- Added hash-mode: Dogechain.info Wallet
- Added hash-mode: Domain Cached Credentials 2 (DCC2), MS Cache 2, (NT)
- Added hash-mode: Domain Cached Credentials (DCC), MS Cache (NT)
- Added hash-mode: Empire CMS (Admin password)
- Added hash-mode: ENCsecurity Datavault (MD5/keychain)
- Added hash-mode: ENCsecurity Datavault (MD5/no keychain)
- Added hash-mode: ENCsecurity Datavault (PBKDF2/keychain)
- Added hash-mode: ENCsecurity Datavault (PBKDF2/no keychain)
- Added hash-mode: GPG (AES-128/AES-256 (SHA-256($pass)))
- Added hash-mode: GPG (AES-128/AES-256 (SHA-512($pass)))
- Added hash-mode: GPG (CAST5 (SHA-1($pass)))
- Added hash-mode: IPMI2 RAKP HMAC-MD5
- Added hash-mode: Kerberos 5, etype 17, AS-REP
- Added hash-mode: Kerberos 5, etype 18, AS-REP
- Added hash-mode: Kremlin Encrypt 3.0 w/NewDES
- Added hash-mode: mega.nz password-protected link (PBKDF2-HMAC-SHA512)
- Added hash-mode: MetaMask Mobile Wallet
- Added hash-mode: MetaMask Wallet (short hash, plaintext check)
- Added hash-mode: Microsoft Online Account (PBKDF2-HMAC-SHA256 + AES256)
- Added hash-mode: MS SNTP
- Added hash-mode: NetIQ SSPR (MD5)
- Added hash-mode: NetIQ SSPR (PBKDF2WithHmacSHA1)
- Added hash-mode: NetIQ SSPR (PBKDF2WithHmacSHA256)
- Added hash-mode: NetIQ SSPR (PBKDF2WithHmacSHA512)
- Added hash-mode: NetIQ SSPR (SHA1)
- Added hash-mode: NetIQ SSPR (SHA-1 with Salt)
- Added hash-mode: NetIQ SSPR (SHA-256 with Salt)
- Added hash-mode: NetIQ SSPR (SHA-512 with Salt)
- Added hash-mode: PDF 1.3 - 1.6 (Acrobat 4 - 8) w/ RC4-40
- Added hash-mode: Perl Mojolicious session cookie (HMAC-SHA256, >= v9.19)
- Added hash-mode: QNX 7 /etc/shadow (SHA512)
- Added hash-mode: RACF KDFAES
- Added hash-mode: RC4 104-bit DropN
- Added hash-mode: RC4 40-bit DropN
- Added hash-mode: RC4 72-bit DropN
- Added hash-mode: RSA Security Analytics / NetWitness (sha256)
- Added hash-mode: SecureCRT MasterPassphrase v2
- Added hash-mode: Veeam VBK
- Added hash-mode: WBB4 (Woltlab Burning Board) Plugin [bcrypt(bcrypt($pass))]

## Constructs

- Added hash-mode: bcrypt(sha256($pass)) / bcryptsha256
- Added hash-mode: bcrypt-sha256 v2 bcrypt(HMAC-SHA256($pass))
- Added hash-mode: md5($salt1.$pass.$salt2)
- Added hash-mode: md5($salt1.sha1($salt2.$pass))
- Added hash-mode: md5($salt1.strtoupper(md5($salt2.$pass)))
- Added hash-mode: md5($salt.md5($pass).$salt)
- Added hash-mode: md5(md5($pass.$salt))
- Added hash-mode: md5(md5($salt).md5(md5($pass)))
- Added hash-mode: md5(md5(md5($pass)).$salt)
- Added hash-mode: md5(md5(md5($pass).$salt1).$salt2)
- Added hash-mode: md5(md5(md5($pass.$salt1)).$salt2)
- Added hash-mode: md5(sha1($pass.$salt))
- Added hash-mode: md5(sha1($salt.$pass))
- Added hash-mode: md5(sha1(md5($pass)))
- Added hash-mode: sha256(sha256($pass.$salt))
- Added hash-mode: sha512(sha512($pass).$salt)
- Added hash-mode: sha512(sha512_bin($pass).$salt)

## Backends

- New: HIP (an OpenCL alternative for AMD on both Linux and Windows)
- New: Metal (an OpenCL alternative on macOS, supporting both ARM and x64)
- New: Microsoft OpenCL D3D12 platform
- Improved handling of devices with unified memory (iGPU, APU, etc.)
- Switched all asynchronous and non-blocking calls to synchronous and blocking ones. This prevents race conditions, such as the one fixed in the previous commit, without any performance impact.
- Revisited the Apple OpenCL 2 GiB bug (still present)
- Added support for longer OpenCL device names
- New: Hardware Monitor support for AMD GPUs via sysfs (Linux)
- New: Hardware Monitor support for CPUs via sysfs (Linux)
- New: Hardware Monitor support for Intel GPUs via sysfs (Linux)

##
## Features
##

## Assimilation Bridge

- Added Assimilation Bridge (See forum post or added documentation for more details)
- Virtual Devices
- Embedded Python interpreter
- Comprehensive documentation

## Docker Support

- Add Dockerfiles for binary compilation
- Compiles both Linux and Windows binaries
- Used to produce official releases for hashcat.net
- Includes Dockerfiles for Ubuntu 20.04 and Arch-based environments
- Supports optional user patches via docker/patches/

## User Features

- Increased custom charset count from 4 to 8
- Added XML support for KeePass keyfiles
- Added true UTF-8 password support for RAR3
- Added JSON format support for backend-info, hash-info, status-screen, progress-info, speed-info, and outfiles
- Added --increment-inverse option
- Added new charset files: Romanian, Russian
- Added support for using --debug-mode in attack-mode 9 (Association Attack)
- Added display of password length mininum and maximum in the Kernel.Feature status line
- Added new feature (-Y) that creates N virtual instances for each device in your system at the cost of N times the device memory consumption
- Added new feature (--bypass-delay / --bypass-threshold) that skips an attack if it does not produce enough cracks per $timeframe
- Added options --benchmark-min and --benchmark-max to set a hash-mode range to be used during the benchmark
- Added option --total-candidates to provide the total candidate count for an attack instead of the internal "--keyspace" value
- Added option --backend-devices-keepfree to configure X percentage of device memory available to keep free
- Added options --benchmark-min and --benchmark-max to set a hash-mode range to be used during the benchmark
- Added the 'edge' test type to test.pl and the corresponding tool test_edge.sh

## Developer Features

- Added support to the tokenizer for multiple signatures per hash mode
- Added option OPTS_TYPE_THREAD_MULTI_DISABLE: allows plugin developers to disable scaling the candidate batch size based on device thread count. Useful for very slow algorithms that parallelize differently
- Added options OPTI_TYPE_SLOW_HASH_DIMY_INIT/LOOP/COMP: enables 2D kernel launches for slow hashes. Y dimension must be set via salt->salt_dimy
- Removed deprecated attribute: OPTS_TYPE_MAXIMUM_ACCEL
- Added KERNEL_FA: a new kernel macro to inform the JIT compiler of thread limits for thread-optimized kernels
- Introduced hashes_init_stage5(), which now calls module_extra_tmp_size(). Self-test hash is fully initialized at this stage
- Improved kernel compiler behavior using hints for register spilling, max threads, and warnings for excessive thread counts
- Improved mapping of functions like hc_bytealign_be(), hc_swap32(), etc., using intrinsics or inline assembly to match hardware capabilities
- Re-enabled USE_BITSELECT, USE_ROTATE, and USE_SWIZZLE for OpenCL
- Improved get_opencl_kernel_wgs() to better derive base values from compute runtimes before autotuning
- Removed --device-as-default-execution-space from nvrtc for hiprtc compatibility. Re-added __device__ to DECLSPEC
- Added support for funnelshift and warp shuffle instructions, including runtime detection and kernel disabling if unsupported
- Added autotune attribute "overtune_unfriendly": allows overtuning thread count under specific conditions
- Added next_power_of_two() and moved both next_power_of_two() and previous_power_of_two() into shared utilities
- Removed default 'rocm-smi' call from benchmark_deep.pl to prevent skewed results
- Reduced default runtime in benchmark_deep.pl due to improved benchmark accuracy
- Redesigned "4-buffer" strategy to prevent excessive memory use from naive division by four, especially in high scrypt configs (e.g., 256k:8:1)
- Added CPU SIMD detection at runtime, relevant for bridge plugins
- Status Code: added specific return code for self-test failure (-11)
- Prepared an enum type to represent final return codes
- Added hcmalloc_aligned() and hcfree_aligned()

## Test framework

- Added new offline testing framework (Test-Edge)
- Added GitHub Actions support for automated build scripts
- Completed large-scale test runs with rarely used combinations (e.g., attack-mode 1 with vector size 16 in ChaCha20)
- Fully reworked install_modules.sh script to automate installation of required Perl and Python modules for the test framework
- Removed many outdated and unstable warnings from existing plugins
- Created repositories to maintain discontinued Perl modules used in the testing framework
- Improved unit test for -m 8300: optimized mode now supports longer passwords, domain names, and salts; both modes enforce a domain name limit of 63 characters
- Fixed SNMPv3 unit test to ensure passwords are at least 8 characters, as required by RFC 3414
- Removed PHP from unit tests entirely
- Added workaround for Net::SSLeay on Apple
- Removed Crypt::GCrypt from the test suite
- Updated install_modules.sh to use cpanm instead of cpan
- Fixed invalid module_constraints for hash-modes 15000, 11100, 10100, 8700, 7801, 7800, 20712, 14400, 21100, 400, and 3100

## Extraction tools

- Added: apfs2hashcat.py
- Added: bisq2hashcat.py
- Added: bitlocker2hashcat.py
- Added: bitwarden2hashcat.py
- Added: cache2data2hashcat.py
- Added: cryptoloop2hashcat.py
- Added: exodus2hashcat.py
- Added: gitea2hashcat.py
- Added: keybag2hashcat.py
- Added: kremlin2hashcat.py
- Added: lastpass2hashcat.py
- Added: luks2hashcat.py
- Added: metamask2hashcat.py
- Added: radmin3_to_hashcat.pl
- Added: shiro1-to-hashcat.py
- Added: truecrypt2hashcat.py
- Added: veeamvbk2hashcat.py
- Added: veracrypt2hashcat.py
- Added: virtualbox2hashcat.py
- Added: vmwarevmx2hashcat.py
- Promoted extraction tools to be installed alongside hashcat when installing system-wide

##
## Improvements
##

## Feature Improvements

- Refactored internal code for Hardware Monitor
- Improved accuracy of benchmark mode (-b)
- Gracefully handle corrupted .gz archives
- Changed benchmark mask to prevent failures during UTF-8 to UTF-16 conversion
- Implemented missing XZ file seeking functionality
- Added workaround to eliminate internal runtime memory leaks

## Rule Engine Improvements

- Added support for character class rules
- Improved many existing rules by removing duplicates and adding new operators, including character class support
- Fixed existing rules to work correctly with the new Purge (@) rule handling
- Updated -j / -k validation to check for ':' instead of NULL
- Optimized the following rule files: generated.rule, generated2.rule, dive.rule, T0XlCv2.rule, T0XlC-insert_00-99_1950-2050_toprules_0_F.rule, T0XlC-insert_space_and_special_0_F.rule, d3ad0ne.rule
- Added new rule files: T0XlC_3_rule.rule, T0XlC_insert_HTML_entities_0_Z.rule, T0XlCv2.rule, stacking58.rule, top10_2025.rule

## Existing Plugin and Crypto Library Improvements

- Improved ASN.1 checks for RSA/DSA/EC/OpenSSH private key modules (22911, 22921, 22931, 22941, 22951)
- AuthMe: updated token length
- VeraCrypt and TrueCrypt: general code cleanup with backported fixes and improvements across both legacy and new modes
- Bitwarden: increased iteration limit
- NSEC3: added support for salt length up to the standard limit; updated max domain and salt lengths in pure kernel to 256
- GOST R 34.11-94: fixed false negative in optimized mode in attack mode 3 for passwords of length 16 or 32
- Fixed hash encoding issue in WinZip format
- Renamed multiple CAST cipher defines to match C++ language naming conventions
- Recommended using --keep-guessing for -m 20510
- Fixed false positive and false negative in -m 21800 multihash mode, where only the first hash was marked as cracked regardless of which was actually cracked (affected only beta versions)
- Fixed encoder bugs in hash-modes 29920 and 29940, where the output string was not properly terminated
- Electrum: added support to detect more private key prefixes
- Metamask: added support for dynamic iteration counts
- TOTP: added support for multiple codes
- Blake2: removed redundant casts and corrected parameter types for FINAL value
- VeraCrypt: set minimum loop count to 250 for all modes with PIM brute-force support
- Fixed thread count issue in -m 10700 on NVIDIA OpenCL (4 bytes were lost per thread for unknown reasons)
- Fixed missing entries in switch_buffer_by_offset_8x4_le_S()
- Applied OPTS_TYPE_NATIVE_THREADS tuning to many hash modes after benchmarking: 770x, 780x, 8900, 9000, 9300, 14000, 14900, 15700, 22700, 2300x, 23900, 24000, 25000, 25100, 25200, 26700, 26800, 26900, 27300, 27700, 28200, 29800, 70000, 70100, 70200, 72000, 73000
- Applied OPTI_TYPE_REGISTER_LIMIT tuning to many hash modes after benchmarking: 20, 1300, 1500, 1800, 6400, 6800, 7700, 7900, 8200, 8300, 10700, 12400, 1375x, 1376x, 14800, 14900, 15600, 16300, 16600, 16700, 16900, 18100, 20800, 21300, 21500, 21700, 22300, 2292x, 2294x, 23400, 23600, 23800, 24200, 24420, 25500, 25900, 26000, 26100, 26600, 26700, 26800, 26900, 27300, 27400, 27500, 27600, 2945x, 2946x
- Improved shared memory handling in -m 10700: removed the 256-thread hard limit and now adapt based on device memory pool
- Check UnpackSize to reduce false positives in hc_decompress_rar
- Added support for zero-length salts in Electrum $4 and $5 formats
- Switched Metamask hashmodes to use TOKEN_ATTR_OPTIONAL_ROUNDS
- Added new hashmodes 26620 and 26630 with dynamic iteration counts
- Added new function count_bits_32() in inc_common.cl
- Simplified module_hash_decode() in -m 32700
- Removed OPTS_TYPE_HASH_COPY from -m 19210
- Fixed missing AES_GCM_decrypt implementation in inc_cipher_aes-gcm.cl
- Optimized same-salt cracking in -m 7400 and -m 500 by marking them as compatible with OPTS_TYPE_DEEP_COMP_KERNEL
- Reduced false positive reports in -m 21800
- WPA: allow users to override nonce_error_corrections even if message_pair suggests otherwise
- Updated Flask session payload max length to 2047
- Added IV support to LastPass hash-mode (-m 6800)
- VeraCrypt: added support for keyfiles and keyfile cascades

## User Interaction Improvements

- Added additional backend information with the new -II parameter
- Set --benchmark to true automatically when --benchmark-all is used
- Improved error message when using --skip/--limit with an invalid attack configuration
- Display the number of base words hashcat expects
- Improved error messages for module_ctx_t validation failures, especially when outdated plugin interfaces are detected
- Added a warning when the user sets a thread count higher than what the runtime recommends (based on available registers and shared memory)
- Silenced selected warning lines when using --quiet
- Added immediate check for --outfile-check-timer
- Adjusted performance warnings for clarity
- Added FAQ link to potfile-related messages
- Increased debug-file flush frequency
- Added [b], [c], and [f] functionality in pause state
- Display all autotune setting changes during benchmark (-b)
- Added support to colorize cracked hashes using --color-cracked
- Display selected warnings only if both --quiet and --machine-readable are false
- Improved salt length reporting in hashconfig output
- Added parameter validation for invalid option combinations:
    - Do not allow --stdout with --slow-candidates
    - Do not allow --show with --restore
    - Do not allow --benchmark with --backend-info
- Display max password length in Kernel.Feature status
- Implemented Windows system information display using -II
- User Options: assigned -H to --hash-info; added detailed output using -HH
- Notify users on startup when backend runtimes and devices are initialized
- Implemented feature request to show non-default session names early at startup
- Warn users that the potfile will not be used during Association Attack mode

##
## Performance
##

- Fully reworked the autotune engine
- Fully reworked scrypt and scrypt-based hash modes (e.g., MultiBit, Electrum)
- Significant speed boost for all QNX hash modes
- Improved performance for all OpenSSH private key modes by reducing the maximum supported password length from 256 to 128
- Reduced JIT compile time by lowering the required C++ standard (C++ is not used)
- Reduced JIT compile time by enabling multithreaded JIT compilation
- DEScrypt Kernel (1500): Improved performance from 950MH/s to 2200MH/s (RX6900XT) on HIP backend by workaround invalid compile time optimizer

## Tuning Database

- Added plugin hook support to generate tuning-db entries on the fly at startup, allowing developers to predict optimal values for kernel-accel and TMTO based on device attributes instead of relying on autotuning
- Updated the list of consumer, mobile, and professional NVIDIA, AMD, and Intel GPUs
- Updated existing vector datatype mappings for NVIDIA
- Added vector datatype support for AMD (previously scalar only)
- Added -m 99999 to Modules_default.hctune

## Memory Management

- Refactored memory management for compute device and host from scratch
- Added automatic downtune support
- Improved free memory detection using low-level API

##
## Bugs
##

- Added verification of token buffer length when using TOKEN_ATTR_FIXED_LENGTH
- Added recovery from rare non-fatal file locking problems
- Added workaround for false positives on 22931
- Fix missing check for -j and -k before writing hashcat.dictstat2 which can lead to false negatives
- Fixed a bug in all SCRYPT-based hash modes with Apple Metal
- Fixed buffer overflow on module_26600.c / module_hash_encode()
- Fixed bug in --stdout that caused certain rules to malfunction
- Fixed bug in --stdout when multiple computing devices are active
- Fixed bug in 18400 module_hash_encode
- Fixed bug in 23800/unrar with Apple Silicon
- Fixed bug in 26900 module_hash_encode
- Fixed bug in 29600 module OPTS_TYPE setting
- Fixed bug in 32600 by adding missing module_jit_build_options
- Fixed bug in 34000 with Metal and OpenCL on Apple
- Fixed bug in Hardware Monitor: prevent disable if ADL fail
- Fixed bug in grep out-of-memory workaround on Unit Test
- Fixed bug in inc_rp_optimized.cl on Apple Intel with Metal
- Fixed bug in input_tokenizer when TOKEN_ATTR_FIXED_LENGTH is used and refactor modules
- Fixed bug in module_constraints and kernel for hash-mode 7800
- Fixed bug in module_constraints and kernel for hash-mode 7801
- Fixed bugs in 3090x kernels which produced false positives
- Fixed bugs in 34100 kernel and inc_luks_aes.cl on Apple Silicon
- Fixed build failed for 10700 optimized with Apple Metal
- Fixed build failed for 13772 and 13773 with Apple Metal
- Fixed build failed for 18400 with Apple Metal
- Fixed build failed for 18600 with Apple Metal
- Fixed build failed for 4410 with vector width > 1
- Fixed build failed for aarch64 (es: rpi)
- Fixed build failure for almost all hash modes that make use of hc_swap64 and/or hc_swap64_S with Apple Metal / Apple Silicon
- Fixed clang identification in src/Makefile
- Fixed debug mode 5 by adding the missing colon between original-word and finding-rule
- Fixed display problem of the "Optimizers applied" list for algorithms using OPTI_TYPE_SLOW_HASH_SIMD_INIT2 and/or OPTI_TYPE_SLOW_HASH_SIMD_LOOP2
- Fixed incompatible pointer types (salt1 and salt2 buf) in 3730 a3 kernel
- Fixed incorrect plaintext check for 25400 and 26610. Increased plaintext check to 32 bytes to prevent false positives.
- Fixed keys extraction in luks2hashcat - now extracts all active keys
- Fixed keys extraction in luks2hashcat - now extracts all LUKS2 keyslots
- Fixed maximum password length in module/test_module of hash-mode 2400
- Fixed maximum password length in module/test_module of hash-mode 2410
- Fixed memory leaks in tuning_db_init in tuningdb.c
- Fixed minimum password length in module of hash-mode 28200
- Fixed minimum password length in module of hash-mode 29800
- Fixed out-of-boundary read when a fast hash defines a kernel_loops_min value higher than the amplifiers provided by the user
- Fixed perl deprecation warnings on test module for 30906
- Fixed race condition in selftest_init on OpenCL with non-blocking write
- Fixed stack buffer overflow in PKZIP modules (17200, 17210, 17220, 17225, 17230)
- Fixed vector datatypes usage for HIP
- Fixed -m 14200 when used in -D1 mode
- Fixed unit tests for hash-modes 20011, 20012, and 20013 (DiskCryptor)
- Fixed XZ file seek operation in hc_fseek()
- Fixed syntax error in vmwarevmx2hashcat
- Fixed bug in Hardware Monitor: prevent disabling if ADL fails
- Fixed bug in --stdout when multiple computing devices are active
- Fixed race condition in selftest_init on OpenCL with non-blocking write
- Fixed division by zero bug in fast hashes caused by hashes->st_salts_buf->salt_iter not being used
- Fixed socket leak in brain_server()
- Fixed incorrect comparison result in sort_pot_orig_line()
- Fixed incorrect comparison result in sort_by_src_len()
- Fixed host buffer overflow when copying rules from host to device
- Handle signed/unsigned PDF permission P value for all PDF hash-modes
- Improved OpenCL function declaration bindings to prevent crashes
- Improve ASN.1 check for RSA/DSA/EC/OpenSSH Private Keys modules (22911, 22921, 22931, 22941, 22951)
- Prevent Hashcat from hanging by checking during startup whether the output file is a named pipe
- Skip chained generated rules that exceed the maximum number of function calls
- Updated benchmark_deep.pl to use ?a?a?a?a?a?a?a instead of ?b?b?b?b?b?b?b to avoid UTF-8 to UTF-16 conversion issues
- Return an error for unsupported arbitrary seek operations
- Solved TODOs in hc_fstat()/filehandling
- Prevented NULL dereference in read_restore() via hcmalloc
- Properly parse maskfiles and escaped question marks

##
## Technical
##

- AMD Driver: Updated requirements for AMD Windows drivers to "AMD Adrenalin Edition" (23.7.2 or later) and "AMD HIP SDK" (23.Q3 or later)
- Alias Devices: Prevents hashcat, when started with x86_64 emulation on Apple Silicon, from showing the Apple M1 OpenCL CPU as an alias for the Apple M1 Metal GPU
- Apple Driver: Automatically enable GPU support on Apple OpenCL instead of CPU support
- Apple Driver: Updated requirements to use Apple OpenCL API to macOS 13.0
- Backend Checks: Describe workaround in error message when detecting more than 64 backend devices
- Backend Info: Added --machine-readable format
- Backend: Added workaround to get rid of internal runtimes memory leaks
- Backend: Splitting backend_ctx_devices_init into smaller runtime-specific functions
- Backend: Updated OpenCL/CUDA/HIP/Metal API's
- Backend: Updated filename chksum format to prevent invalid cache on Apple Silicon when switching arch
- Backend: Updated references of hcmalloc_aligned and hcfree_aligned to the new memory defined functions
- Benchmark: Updated benchmark_deep.pl with OS detection, Apple cache cleaning way and other minor changes
- Brain: Added sanity check and corresponding error message for invalid --brain-port values
- Bridge: Updated references of hcmalloc_aligned and hcfree_aligned to the new memory defined functions
- Building: Support building windows binaries on macOS using MinGW
- Debug: Added -g to build_options if DEBUG >= 1 (only with HIP and OpenCL)
- Dependencies: Added sse2neon v1.8.0 (commit 658eeac)
- Dependencies: Updated LZMA SDK to 24.09
- Dependencies: Updated zlib to 1.3.1
- Dependencies: Updated OpenCL-Headers to v2024.10.24 (commit 265df85)
- Dependencies: Updated unrar source to 6.2.7
- Dependencies: Updated xxHash to 0.8.3 (commit 50f4226)
- Documents: Renamed status_code.txt in exit_status_code.txt and added device_status_code.txt
- Documents: Updated BUILD.md and added BUILD_macOS.md (containing instructions for building windows binaries on macOS)
- HIP Backend: Avoid deprecated functions
- Hardware Monitor: Added current power consumption indicator (only supported on Apple)
- Hardware Monitor: Added hm_get_power_* primitives (only supported on Apple)
- Hardware Monitor: Renamed src/ext_iokit.c to src/ext_iokit.m
- Hardware Monitor: Splitting hwmon_ctx_init function into smaller library-specific functions
- Hardware Monitor: avoid sprintf in src/ext_iokit.c
- Hash-Info: show more details using -HH
- Help: show supported hash-modes only with -hh
- Kernel: Renamed multiple defines in CAST cipher to fit expected naming convention of the C++ language standard
- Makefile: Added -lIOReport into LFLAGS_NATIVE for Apple and moved ext_iokit into OBJS_METAL
- Makefile: prevent make failure with Apple Silicon in case of partial rebuild
- Makefile: updated MACOSX_DEPLOYMENT_TARGET to 15.0
- MetaMask: update extraction tool to support MetaMask Mobile wallets
- Memory: added hc_alloc_aligned and hc_free_aligned
- Memory: renamed hcmalloc_aligned and hcfree_aligned to hcmalloc_bridge_aligned and hcfree_bridge_aligned
- Metal Backend: Added hc_mtlFinish
- Metal Backend: added support to 2D/3D Compute
- Metal Backend: added workaround to prevent 'Infinite Loop' bug when build kernels
- Metal Backend: added workaround to set the true Processor value in Metal devices on Apple Intel
- Metal Backend: allow use of devices with Metal if runtime version is >= 200
- Metal Backend: define USE_BITSELECT and USE_SWIZZLE for Apple Intel
- Metal Backend: disable Metal devices only if at least one OpenCL device is active
- Metal Backend: improved compute workloads calculation
- Metal Backend: improved memory management
- Metal Backend: fixed memory leaks
- Metal Backend: parallelize pipeline state object (PSO) compilation internally
- Metal Backend: Updated hc_mtlCreateBuffer
- Modules: Added OPTS_TYPE_PT_BASE58 in 28501 28502 28503 28504 28505 28506 30901 30902 30903 30904 30905 30906
- Modules: Added module_unstable_warning for 22500, update module_unstable_warning for 10700
- Modules: Added support for non-zero IVs for -m 6800 (Lastpass). Also added `tools/lastpass2hashcat.py`
- Modules: Check UnpackSize to raise false positive with hc_decompress_rar
- Modules: Updated module_unstable_warning
- Open Document Format: Added support for small documents with content length < 1024
- OpenCL Backend: added workaround to set device_available_memory from CUDA/HIP alias device
- OpenCL Backend: Added hc_clCreateBuffer wrapper: hc_clCreateBuffer_pre
- OpenCL Backend: using CL_TRUE with all hc_clEnqueueWriteBuffer() call
- OpenCL Backend: using HC_OCL_CREATEBUFFER macro for buffer allocation and openclMemoryFlags array to configure the memory flags
- Rules: Add support to character class rules
- Rules: Rename best64.rule to best66.rule and remove the unknown section from it
- Scrypt: Increase buffer sizes in module for hash mode 8900 to allow longer scrypt digests
- SecureCRT MasterPassphrase v2: update module, pure kernels and test unit. Add optimized kernels.
- Selftest: rename selftest function to process_selftest and splitting into 3 smaller functions
- Status Code: Add specific return code for self-test fail (-11)
- Types: Added OPTS_TYPE_PT_BASE58 and removed OPTS_TYPE_MAXIMUM_ACCEL in opts_type
- Unicode: Update UTF-8 to UTF-16 conversion to match RFC 3629
- Unit tests: Updated install_modules.sh, test.pl and some test modules (1000, 16000, 31300, 31500, 31600) by removing Text:Iconv and using Encode instead
- Unit tests: Updated install_modules.sh with Crypt::Argon2
- Unit tests: Updated install_modules.sh with Crypt::Passwd::XS, to test suite works also on Apple (ex: 1800)
- Unit tests: Updated install_modules.sh to remove php and Crypt::GCrypt (updated relative test modules)
- Unit tests: Updated install_modules.sh to use an external module for Digest::BLAKE2, due to maintenance being discontinued
- Unit tests: Updated install_modules.sh to use an external module for Digest::GOST, due to maintenance being discontinued
- Unit tests: Updated install_modules.sh to use an external package for pygost, due to not available anymore
- Unit tests: Updated install_modules.sh to use cpanm instead of cpan
- Unit tests: Updated install_modules.sh with Crypt::Argon2
- Unit tests: Updated install_modules.sh with Crypt::Passwd::XS, to test suite works also on Apple (ex: 1800)
- Unit tests: Updated install_modules.sh, test.pl and some test modules (1000, 16000, 31300, 31500, 31600) by removing Text:Iconv and using Encode instead
- Unit tests: Updated test_edge.sh, introducing clean_cache function and fix for 20510 module
- User Options: Added error message when mixing --username and --show to warn users of exponential delay
- User Options: added --metal-compiler-runtime option
- User Options: assigned -H to --hash-info
- User Options: do not allow --slow-candidates (-S) in benchmark mode
- User Options: limit --bitmap-max value to 31

## Workarounds for 3rd party runtimes and libraries

- Added hc_uint4_t to scrypt to work around an Intel OpenCL alignment bug. Intel's runtime does not support vector datatypes mapped to global memory, though they work with local variables
- Revisited all module_unstable_warning entries across all plugins and synchronized them with the current state of driver versions at release time (most warnings have been removed)
- Improved benchmark stability: hashcat no longer creates and destroys context and command queues for each device when switching hash-modes. This avoids GPU memory leaks with clCreateContext, especially on NVIDIA OpenCL

## Documents

- New: hashcat-assimilation-bridge.md
- New: hashcat-assimilation-bridge-development.md
- New: hashcat-python-plugin-quickstart.md
- New: hashcat-python-plugin-requirements.md
- New: hashcat-python-plugin-development-guide.md
- New: BUILD_macOS.md
- New: BUILD_Docker.md
- New: hashcat-example-hashes.md (auto-generated)
- New: hashcat-help.md (auto-generated)
- New: device_status_code.txt

## Building and Installing

- Refactored GitHub Actions workflow to use matrix strategy for more efficient builds
- Fixed build failure on aarch64 platforms (e.g., Raspberry Pi)
- Updated BUILD instructions for CYGWIN and MSYS2
- Updated build flags on macOS (version placeholder: XX)
- Updated dependencies including UNRAR and OpenCL headers
- Automatically enabled kernel mode debugging for JIT-compiled HIP and OpenCL kernels when hashcat is built with DEBUG=1
- Fixed all compiler warnings for both GCC and Clang (latest versions) in hashcat source; suppressed warnings in third-party libraries
- Fixed compilation error in MSYS2 native shell
- Fixed Clang version detection in src/Makefile
- Changed package script source directory from $HOME/hashcat to current directory
- Removed old iconv patches (now handled by CMake)
- Fixed bash completion install script
- Updated WSL documentation to use CMake for win-iconv setup
- Fixed MinGW printf format issues
- Fixed compilation on newer FreeBSD versions
- Included winsock2.h instead of winsock.h
- Silenced Clang32/64 warnings related to signed/unsigned comparisons
- Fixed stdcall-related warnings under Clang32/64

##
## Plans for upcoming releases
##
- Add more assimilation bridge plugins
- Rewrite testing framework from Perl to Python

* changes v6.2.5 -> v6.2.6

##
## Algorithms
##

- Added hash-mode: Amazon AWS4-HMAC-SHA256
- Added hash-mode: Bitcoin WIF private key (P2PKH)
- Added hash-mode: Bitcoin WIF private key (P2SH(P2WPKH))
- Added hash-mode: Bitcoin WIF private key (P2WPKH, Bech32)
- Added hash-mode: BLAKE2b-512($pass.$salt)
- Added hash-mode: BLAKE2b-512($salt.$pass)
- Added hash-mode: DPAPI masterkey file v1 (context 3)
- Added hash-mode: DPAPI masterkey file v2 (context 3)
- Added hash-mode: Exodus Desktop Wallet (scrypt)
- Added hash-mode: Flask session cookie
- Added hash-mode: KeePass 1 (AES/Twofish) and KeePass 2 (AES) - keyfile only mode
- Added hash-mode: Kerberos 5, etype 17, DB
- Added hash-mode: Kerberos 5, etype 18, DB
- Added hash-mode: PostgreSQL SCRAM-SHA-256
- Added hash-mode: Radmin3
- Added hash-mode: Teamspeak 3 (channel hash)
- Added hash-mode: Terra Station Wallet (AES256-CBC(PBKDF2($pass)))
- Added hash-mode: bcrypt(sha512($pass)) / bcryptsha512
- Added hash-mode: md5(sha1($pass).$salt)
- Added hash-mode: sha1($salt.sha1(utf16le($username).':'.utf16le($pass)))
- Added hash-mode: sha256($salt.sha256_bin($pass))

##
## Features
##

- Added new backend support for Metal, the OpenCL replacement API on Apple
- Added support for building universal macOS binary on Apple Silicon
- Added support for using --debug-mode in attack-mode 9 (association attack)
- Added hex encoding format for --separator option
- Added password candidates range to --status-json output
- Added parameter to Bitwarden mode for second iteration count
- Added support for using 'John the Ripper' hash format with hash-type 13100
- Added support for using 'John the Ripper' hash format with hash-type 18200
- Added hash extraction scripts from the tools folder to beta/release versions
- Added user advice if a hash throws 'token length exception'
- Added tunings/ folder in order to replace hashcat.hctune. Configuration files with *.hctune suffix are automatically load on startup
- Added link to 'ubernoob' documentation when no parameters are supplied on Windows

##
## Bugs
##

- Fixed accepted salt length by PKCS#8 Private Keys modules
- Fixed autodetect memory allocation for temporary hashes for LUKS v1 (legacy) in --identify mode
- Fixed backend active devices checks
- Fixed building error on Raspberry Pi
- Fixed display problem of incorrect negative values in case of large numbers
- Fixed display problem of the "Optimizers applied" list for algorithms using Register-Limit
- Fixed example password output of --hash-info: force uppercase if OPTS_TYPE_PT_UPPER is set
- Fixed false negative on hash-type 27800 if using vector width greater than 1 and -a 3
- Fixed false negative on hash-types 4510 and 4710 for hashes with long salts
- Fixed false negative on hash-types 8900, 15700, 22700, 27700 and 28200 if using the HIP backend
- Fixed false negative on Unit Test in case of out-of-memory with grep in single mode
- Fixed false negative on Unit Test with hash-type 25400
- Fixed functional error when nonce-error-corrections that were set on the command line in hash-mode 22000/22001 were not accepted
- Fixed handling of devices in benchmark mode for "kernel build error". Instead of canceling, skip the device and move on to the next
- Fixed handling of password candidates that are shorter than the minimum password length in Association Attack
- Fixed invalid handling of keyfiles in Keepass if transf_random_seed doesn't change
- Fixed memory leak in CPU rule engine
- Fixed method of how OPTS_TYPE_AUX* kernels are called in an association attack, for example in WPA/WPA2 kernel
- Fixed missing option flag OPTS_TYPE_SUGGEST_KG for hash-mode 11600 to inform the user about possible false positives in this mode
- Fixed optimized (-O) candidate generation with --stdout and -a 7
- Fixed password limit in optimized kernel for hash-mode 10700
- Fixed password reassembling function reporting an incorrect candidate in some cases when the correct candidate has zero length
- Fixed undefined function call to hc_byte_perm_S() in hash-mode 17010 on non-CUDA compute devices
- Fixed unit test early exit on luks test file download/extract failure
- Fixed unit test false negative if there are spaces in the filesystem path to hashcat
- Fixed unit test salt-max in case of optimized kernel, with hash-type 22 and 23
- Fixed usage of --rule-right (-k) in -a 7 with optimized (-O) kernels
- Fixed wordlist handling in -m 3000 when candidate passwords use the $HEX[...] syntax

##
## Technical
##

- AMD Driver: Updated requirements for AMD Linux drivers to "AMDGPU" (21.50 or later) and "ROCm" (5.0 or later)
- AMD Driver: Updated requirements for AMD Windows drivers to "AMD Adrenalin Edition" (Adrenalin 22.5.1 exactly)
- Association Attack: Enable module specific pw_min and pw_max settings to avoid false positives in -a 9 attack-mode
- Autotune: Added error handling. By default skipping device on error, with --force using accel/loops/threads min values instead
- Backend: improved management of systems with multiple OpenCL platforms
- Backend Info: Added folder_config info to output
- Backend Info: Added generic system info to output (must be completed on Windows side)
- Backend Info: Added local memory size to output
- Backend: with kernel build options, switch from -I to -D INCLUDE_PATH, in order to support Apple Metal runtime
- Command Line: Disallow combinations of some options. for instance, using -t in -a 0 mode
- CUDA Backend: moved functions to ext_cuda.c/ext_nvrtc.c and includes to ext_cuda.h/ext_nvrtc.h
- Debug Rules: Set --debug-file to $session.debugfile if --debug-mode was set by the user and --debug-file was not set
- Hardware Monitor: Add support for GPU device utilization readings using iokit on Apple Silicon (OpenCL and Metal)
- Hash Info: show more information (Updated Hash-Format. Added Autodetect, Self-Test, Potfile and Plaintext encoding)
- HIP Backend: moved functions to ext_hip.c/ext_hiprtc.c and includes to ext_hip.h/ext_hiprtc.h
- HIP Backend: removed unused functions from hiprtc to workaround missing function symbols on windows dll
- Kernels: Refactored standard kernel declaration to use a structure holding u32/u64 attributes to reduce the number of attributes
- Kernels: Refactored standard kernel includes, KERN_ATTR macros and RC4 cipher functions, in order to support Apple Metal runtime
- Kernels: Set the default Address Space Qualifier for any pointer, in order to support Apple Metal runtime
- Logfile: Write per-session "recovered new" value to logfile
- Makefile: updated MACOSX_DEPLOYMENT_TARGET to 10.15 and removed OpenCL framework from LFLAGS_NATIVE on MacOS
- Metal Runtime: added support for vectors up to 4
- Modules: Added suffix *legacy* to old TrueCrypt modules (6211-6243)
- Modules: Added suffix *legacy* to old VeraCrypt modules (13711-13783)
- Modules: Added support of a custom charset setting for benchmarks to the module interface
- Modules: New LUKS v1 modules (29511-29543) which do not use `module_hash_binary_parse` to get data from containers anymore (use new tool `tools/luks2hashcat.py`)
- Modules: New TrueCrypt modules (29311-29343) which do not use `module_hash_binary_parse` to get data from containers anymore (use new tool `tools/truecrypt2hashcat.py`)
- Modules: New VeraCrypt modules (29411-29483) which do not use `module_hash_binary_parse` to get data from containers anymore (use new tool `tools/veracrypt2hashcat.py`)
- Modules: Renamed old LUKS module into LUKS v1 and added suffix *legacy* (14600)
- OpenCL Backend: added workaround to make optimized kernels work on Apple Silicon
- OpenCL Backend: moved functions to ext_OpenCL.c and includes to ext_OpenCL.h
- OpenCL Backend: show device_type in device list info on Apple Silicon
- OpenCL Kernel: Set native_threads to 32 on Apple GPU's for various hash-modes
- OpenCL Runtime: Added support to use Apple Silicon compute devices
- OpenCL Runtime: Add some unstable warnings detected on macOS
- OpenCL Runtime: Set default device-type to GPU with Apple Silicon compute devices
- Restore: Restore timer is decreased from 60 seconds to 1 second, but only updates if there's actually a change compared to previous data written to restore file
- Rules: Add new rulesets from T0XlC: T0XlCv2, T0XlC_3_rule, T0XlC_insert_HTLM_entities_0_Z
- Rules: Add support to include source wordlist in debugging format
- Rules: Update hand-written rulesets to covers years up to 2029
- Status code: updated negative status code (added kernel create failure and resync)
- Status code: updated negative status code, useful in Unit tests engine (test.sh)
- Terminal: Increased size of hash name column in `--help` and `--identify` options
- Terminal: Limit output length of example hash in --example-hash mode to 200. Use --mach to see full example hash
- Terminal: show empty OpenCL platforms only in backend information mode
- Tuning Database: Added a warning if a module implements module_extra_tuningdb_block but the installed computing device is not found
- Unit tests: added -r (--runtime) option
- Unit tests: handle negative status code, skip deprecated hash-types, skip hash-types with known perl modules issues, updated output
- Unit tests: Updated test.sh to set default device-type to CPU with Apple Intel and added -f (--force) option
- Usage Screen: On windows console, wait for any keypress if usage_mini_print() is used
- User Options: Add new module function module_hash_decode_postprocess() to override hash specific configurations from command line
- User Options: Change --backend-info/-I option type, from bool to uint
- Workflow: Added basic workflow for GitHub Actions

* changes v6.2.4 -> v6.2.5

##
## Algorithms
##

- Added hash-mode: CRC32C
- Added hash-mode: CRC64Jones
- Added hash-mode: MultiBit Classic .wallet (scrypt)
- Added hash-mode: MurmurHash3
- Added hash-mode: Windows Hello PIN/Password

##
## Performance
##

- PDF Kernel (10700): Improved performance on AMD GPUs by using shared memory for the scratch buffer and disabled inlining to save spilling

##
## Bugs
##

- Fixed divide by zero error because backend_ctx->hardware_power_all was not re-inserted after refactoring device_param->hardware_power
- Fixed invalid handling of initialization value for -m 11500
- Fixed invalid progress counter initialization in attack-mode 9 when using --skip or --restore
- Fixed out-of-boundary reads in attack-mode 9 that were caused by a missing work item limit in the refactored autotune engine
- Fixed out-of-boundary reads in hash-mode 22400 (AES Crypt) kernel
- Fixed strategy for eliminating hashes with zero length in LM when multiple hashes contain a zero hash

##
## Technical
##

- AMD Driver: Updated requirements for AMD Linux drivers to ROCm 4.5 or later due to new HIP interface
- Backend devices: In -S mode, limit the number of workitems so that no more than 2GB of host memory is required per backend device
- Backend devices: In non -S mode, limit the number of workitems so that no more than 4GB of host memory is required per backend device
- Backend types: The default filter for the device types is now set so that only the GPU is used, except for APPLE, where we set CPU
- Benchmark: Update benchmark_deep.pl with new hash modes added (also new hash modes which were added with v6.2.3)
- Electrum Wallet: Added new entropy-based check to test whether the decryption was successful or not
- Module Optimizers: Added OPTS_TYPE_MAXIMUM_THREADS to deactivate the else branch route in the section to find -T before compilation
- Makefile: Added wildcard include src/modules/module_*.mk directive so that plugin developers can add 3rd party libraries for their plugins
- Rejects: Disabled checking of the minimum and maximum length of the password candidate in attack-mode 9 because they are incompatible
- POCL: Added a workaround for an issue in POCL where a quote character is used as a part of the path itself which is passed as a path for the -I option
- Device Threads: The default maximum device thread number has been reduced from 1024 to 256, users can still overwrite with the -T option
- Tuning-DB: Add missing entries for -m 25600 and -m 25800 for CPU cracking
- OpenCL Backend: added workaround to support Apple Silicon (GPU: M1, M1 Pro, and M1 Max)
- MacOS v10.8+ (PBKDF2-SHA512): Added support for parsing new mac2john hash format directly in the module for -m 7100

* changes v6.2.3 -> v6.2.4

##
## Algorithms
##

- Added hash-mode: SNMPv3 HMAC-MD5-96/HMAC-SHA1-96
- Added hash-mode: SNMPv3 HMAC-MD5-96
- Added hash-mode: SNMPv3 HMAC-SHA1-96
- Added hash-mode: SNMPv3 HMAC-SHA224-128
- Added hash-mode: SNMPv3 HMAC-SHA256-192
- Added hash-mode: SNMPv3 HMAC-SHA384-256
- Added hash-mode: SNMPv3 HMAC-SHA512-384
- Added hash-mode: VirtualBox (PBKDF2-HMAC-SHA256 & AES-128-XTS)
- Added hash-mode: VirtualBox (PBKDF2-HMAC-SHA256 & AES-256-XTS)

##
## Features
##

- Added option --deprecated-check-disable to enable deprecated plugins
- Added option --generate-rules-func-sel to specify a pool of operators that the random rule generator is allowed to pick from
- Added option --multiply-accel-disable (short: -M) to disable multiplying of the kernel accel with the multiprocessor count
- Added rule function '3' to change the case of the first letter after the occurrence of N of character X
- Added support for auto tuning --kernel-threads (-T) at startup
- Added support for HIP version 4.3 or later and removed support for older HIP versions as they are not compatible

##
## Bugs
##

- Fixed broken support for --generate-rules-func-min which was ignored under certain conditions
- Fixed buffer overflow in DPAPI masterkey file v1/v2 module in hash_encode() and hash_decode()
- Fixed buffer overflow in Stargazer Stellar Wallet XLM module in hash_encode() when a hash was cracked
- Fixed false negative in all VeraCrypt hash modes if both conditions are met: 1. Use CPU for cracking and 2. PIM area was used
- Fixed invalid data type in the sha384_hmac_init_vector_128() function that take effect if the vector data type was specified manually
- Fixed out-of-boundary read in input_tokenizer() if the signature in the hash is longer than the length of the plugin's signature constant
- Fixed out-of-boundary read in PKZIP masterkey kernel if the password candidate has length zero
- Fixed out-of-boundary read in the Stuffit5 module in hash_decode()
- Fixed random rule generator option --generate-rules-func-min by fixing switch() case to not select a not existing option group type
- Fixed segfault when a combination of the flags --user and --show is given and a hash was specified directly on the command line
- Fixed syntax check of HAS_VPERM macro in several kernel includes causing invalid error message for AMD GPUs on Windows
- Fixed uninitialized tmps variable in autotune for slow hashes by calling _init and _prepare kernel before calling _loop kernel
- Fixed unintentional overwriting of the --quiet option in benchmark and speed-only mode

##
## Performance
##

- AMD GPUs: Add inline assembly code for md5crypt, sha256crypt, PDF 1.7, 7-Zip, RAR3, Samsung Android and Windows Phone 8+
- AMD GPUs: On the Apple OpenCL platform, we ask for the preferred kernel thread size rather than hard-coding 32
- Backend Interface: Replace most of the blocking Compute API functions with asynchronous ones to improve GPU utilization
- Blake Kernels: Optimize 3/4 BLAKE2B_ROUND() 64-bit rotations with inline assembly hc_byte_perm_S() calls
- Blowfish Kernels: Backport optimizations to reduce bank conflicts from bcrypt to Password Safe v2 and Open Document Format (ODF) 1.1
- ECC secp256k1: The inline assembly code for AMD GPUs has been removed as the latest JIT compilers optimize it with the same efficiency
- HIP Kernels: Enable vector data types for HIP kernels for functionality and performance
- Kernel threads: Use warp size / wavefront size query instead of hard-coded values as the basis for kernel threads
- SCRYPT Kernels: Improve Hashcat.hctune entries for many NV and AMD GPUs for hash mode 8900, 9300, 15700 and 22700
- STDIN: Reduce the performance gap between stdin and wordlist by disabling the --stdin-timeout-abort check after 1000 successful reads
- Tuning Database: Add new module function module_extra_tuningdb_block() to extend hashcat.hctune content from a module

##
## Technical
##

- 7-Zip Hook: Increase the supported data length from 320kb to 8mb
- ADL: Updated support for AMD Display Library to 15.0, updated data types
- AMD Driver: Updated requirements for AMD Linux drivers to ROCm 4.3 or later due to new HIP interface
- AMD Driver: Updated requirements for AMD Windows drivers to Adrenalin 21.2.1 or later due to new ADL library
- Autotune: Add autotune event to inform the user about autotune is starting/stopping on startup
- Backend Interface: Implement gpu_bzero() as a gpu_memset() replacement, since all gpu_memset() operations used 0 as the value
- Backend Interface: Improve the query kernel's dynamic memory size based on DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN instead of BF
- Brain Session: Adds hashconfig-specific opti_type and opts_type parameters to the session calculation to enable cover functions like -O
- Commandline: Fix some memory leaks in case hashcat aborts due to command line parameter settings
- Commandline: Throw an error if the separator specified by the user with the -p option is not exactly 1 byte
- Community: Add link to new hashcat discord to docs/contact.txt
- Constants: Make const char * pointers actually const char * const pointers
- Deprecated Plugins: Add new module function module_deprecated_notice() to mark a plugin as deprecated and to return a free text user notice
- Deprecated Plugins: Marked plugins 2500/2501 and 16800/16801 as deprecated
- Encoding: Truncate password candidates in UTF8 -> UTF16 conversion if it contains an invalid UTF8 byte sequence
- Filehandling: Avoid direct access to some file handles and use internal structures instead
- Filehandling: Use const char for fopen mode to fix -Wwrite-strings warnings
- Hardware Monitor: Added support for OverDrive 7 and 8 based GPUs
- HIP Kernels: Dependency on hip/hip runtime.h has been removed to enable easier integration of the HIP backend under Windows
- Kernel cache: Add kernel threads for hash calculation, which will later be used in the file name of the kernel cache
- Memory Management: Check in OpenCL that CL_DEVICE_HOST_UNIFIED_MEMORY is true and if so, then half of the available memory
- Memory Management: Refactored the code responsible for limiting kernel accel with the goal to avoid low host memory situations
- MetaMask: Added special extraction tool for MetaMask wallets
- NEO driver: Activate the NEO driver (Intel GPU) after it passed several tests in hashcat using the latest NEO driver version
- OpenCL Runtime: The use of clUnloadPlatformCompiler () was disabled after some users received unexpected return codes
- OpenCL Runtime: Workaround for Intel OpenCL runtime: segmentation fault when compiling hc_enc_next() / hc_enc_next_global()
- Potfile: Fixed some leaks in potfile_*_open(), potfile_*_close(), potfile_destroy() and don't assume plain file
- RC4 Kernels: Use improved native thread derivation for RC4-based hash modes 7500, 13100, 18200, 25400
- Shared Memory: Calculate the dynamic memory size of the kernel based on CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN
- Slow kernels: Set some of the slowest kernels to OPTS_TYPE_MP_MULTI_DISABLE to make it easier to handle small word lists
- Status view: Add hash-mode (-m number) in addition to hash-name
- Vendor Discovery: Add "Intel" as a valid vendor name for GPUs on macOS
- MetaMask: Increase the supported data len from 784b to 3136b and set pw-min to 8

* changes v6.2.2 -> v6.2.3

##
## Algorithms
##

- Added hash-mode: AES-128-ECB NOKDF (PT = $salt, key = $pass)
- Added hash-mode: AES-192-ECB NOKDF (PT = $salt, key = $pass)
- Added hash-mode: AES-256-ECB NOKDF (PT = $salt, key = $pass)
- Added hash-mode: iPhone passcode (UID key + System Keybag)
- Added hash-mode: MetaMask Wallet
- Added hash-mode: VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 512 bit  + boot-mode
- Added hash-mode: VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1024 bit + boot-mode
- Added hash-mode: VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1536 bit + boot-mode

##
## Features
##

- Added new backend support for HIP, an OpenCL alternative API for AMD GPUs (similar to CUDA for NVIDIA GPUs)
- Added option --markov-inverse to inverse markov statistics, with the idea of reversing the order of the password candidates
- Added temperature watchdog and fanspeed readings for CPU and GPU on macOS using iokit
- Added temperature watchdog and utilization for CPU on linux using sysfs and procfs
- Added option --increment-inverse to increment masks from right-to-left instead of left-to-right

##
## Bugs
##

- Fixed access violation in hashconfig_destroy if hashcat_ctx_t is only partially initialized
- Fixed 'E' rule in pure kernel mode which was ignoring letters that are in positions that are multiples of 4
- Fixed false negative in hash-mode 15900 (DPAPI masterkey file v2) if password was longer than 64 characters
- Fixed hashcat_ctx leak and refactor module and kernel existence checks
- Fixed integer overflow in Recovered/Time status view column caused by division > 0 but < 1
- Fixed invalid ETA if --limit was specified, several salts are in a hash list and some of them were found in a potfile
- Fixed memory leak in iconv_ctx and iconv_tmp in backend.c
- Fixed missing option to automatically disable kernel cache in -m 25600 and -m 25800
- Fixed out-of-boundary write in slow candidates mode in combinator attack

##
## Improvements
##

- Alias Devices: Show a warning in case the user specifically listed a device to use which in a later step is skipped because it is an alias of another active device
- Backend Information: Added new column showing the PCI Address per CUDA/OpenCL device to easier identify broken cards
- Bcrypt-SHA1/MD5 Kernels: Get rid of local memory use for binary to hex conversion to avoid false negatives on several OpenCL runtimes
- CPU Affinity: Allow mask up to 64 processors in Windows and remove call to SetThreadAffinityMask as SetProcessAffinityMask limits all threads in process
- Debug Rules: HEX-ify rules debug non-rules outputs that contain colons
- KeePass: Increase supported size for KeePass 1 databases from 300kB to 16MB
- Potfile: Disable potfile for hash-mode 99999
- VeraCrypt: Increase password length support for non-boot volumes from 64 to 128
- WPA Kernels: Increased performance by 3.5% for backend devices controlled by CUDA backend

##
## Technical
##

- Autodetect: Limit the number of errors per hash-mode try to 100 to avoid long startup time
- Brain: Add brain_ctx_t to hashcat_ctx_t to enable runtime check if hashcat was compiled with brain support
- File handling: Do not abort on seeing a BOM in input files, just warn and ignore the BOM
- Folders: Do not escape the variable cpath_real to prevent certain OpenCL runtimes from running into an error which do not support escape characters
- Java Object hashCode: Add OPTS_TYPE_SUGGEST_KG as a default option
- LM: Workaround JiT compiler bug in -m 3000 on NV leading to false negatives with large amount of hashes
- OpenCL Runtime: Workaround a problem of the AMD OpenCL GPU driver under macOS which results in false negatives due to changes in the kernel crypto library
- OpenCL Runtime: Workaround JiT crash (SC failed. No reason given.) on macOS by limiting local memory allocations to 32k
- Status View: Include time and duration info when pausing and resuming
- Tests: Changed tests for VeraCrypt from -a 0 to -a 3, because password extension is not available to all shells
- WinZip Kernel: Increase supported data length from 8k to 16mb

* changes v6.2.1 -> v6.2.2

##
## Algorithms
##

- Added hash-mode: bcrypt(md5($pass)) / bcryptmd5
- Added hash-mode: bcrypt(sha1($pass)) / bcryptsha1
- Added hash-mode: FortiGate256 (FortiOS256)
- Added hash-mode: Linux Kernel Crypto API (2.4)
- Added hash-mode: MurmurHash
- Added hash-mode: OpenEdge Progress Encode
- Added hash-mode: md5(utf16le($pass))
- Added hash-mode: sha1(utf16le($pass))
- Added hash-mode: sha256(utf16le($pass))
- Added hash-mode: sha384(utf16le($pass))
- Added hash-mode: sha512(utf16le($pass))
- Added hash-mode: md5(md5(md5($pass)))
- Added hash-mode: sha1(sha1($salt.$pass.$salt))
- Added hash-mode: sha256($salt.sha256($pass))
- Added hash-mode: sha384($pass.$salt)
- Added hash-mode: sha384($salt.$pass)
- Added hash-mode: sha384($salt.utf16le($pass))
- Added hash-mode: sha384(utf16le($pass).$salt)

##
## Features
##

- Autodetect hash-type: performs an automatic analysis of input hash(es), either listing compatible algorithms, or executing the attack (if only one compatible format is found)
- Autodetect hash-type: added option --identify to only perform autodetection of hash-type, without back-end device initialization

##
## Bugs
##

- Fixed error message in -a 9 mode with rules when number of words from wordlist is not in sync with number of unique salts
- Fixed error-handling logic in monitor thread to not return in case of error (disk full, permission error, ...) but to retry instead
- Fixed false negatives with TrueCrypt/VeraCrypt when zip- or gzip-compressed files were used as keyfiles
- Fixed free memory-size output for skipped GPU (both automatic and manual) of --backend-info information screen

##
## Improvements
##

- AES Crypt Plugin: Reduced max password length from 256 to 128 which improved performance by 22%
- CUDA Backend: If --stdout is used, do not warn about missing CUDA SDK
- Folder Management: Add support for XDG Base Directory specification when hashcat is installed using 'make install'
- Hardware Monitor: Add support for GPU device utilization readings from sysfs (AMD on Linux)
- OpenCL Backend: Use CL_DEVICE_BOARD_NAME_AMD instead of CL_DEVICE_NAME for device name (when supported by OpenCL runtime)
- Performance Monitor: Suggest -S to improve cracking performance in specific attack configurations
- RAR3-p (Compressed): Fix workaround in unrar library in AES constant table generation to enable multi-threading support
- RC4 Kernels: Improved performance by 20%+ with new RC4 code in hash-modes Kerberos 5 (etype 23), MS Office (<= 2003) and PDF (<= 1.6)
- Scrypt Kernels: Re-enable use of kernel cache in scrypt-based kernels
- Status Screen: Show currently-running kernel type (pure, optimized) and generator type (host, device)
- UTF8-to-UTF16: Replaced naive UTF8-to-UTF16 conversion with true conversion for RAR3, AES Crypt, MultiBit HD (scrypt) and Umbraco HMAC-SHA1

##
## Technical
##

- Dependencies: Updated LZMA SDK from 19.00 to 21.02 alpha
- Dependencies: Updated xxHash from 0.1.0 to v0.8.0 - Stable XXH3
- Documentation: Update missing documentation in plugin developer guide for OPTS_TYPE_MP_MULTI_DISABLE and OPTS_TYPE_NATIVE_THREADS
- Hashrate: Update inner-loop hashrate prediction to handle new salt_repeats feature and also respect _loop2 kernel runtime
- Kernels: Add RC4 cipher to crypto library with optimized shared memory access pattern which will not cause any bank conflicts if -u <= 32
- Kernels: Add standalone true UTF8-to-UTF16 converter kernel that runs after amplifier. Use OPTS_TYPE_POST_AMP_UTF16LE from plugin
- Kernel Cache: Add module_jit_build_options() string from modules to kernel cache checksum calculation
- Modules: Recategorized HASH_CATEGORY option in various modules
- Modules: Added hash categories HASH_CATEGORY_IMS and HASH_CATEGORY_CRYPTOCURRENCY_WALLET
- Modules: Changed hash category of Python passlib from HASH_CATEGORY_GENERIC_KDF to HASH_CATEGORY_FRAMEWORK
- Unit-Tests: Added missing support for OPTS_TYPE_LOOP_PREPARE, OPTS_TYPE_LOOP_PREPARE2 and salt_repeats in self-test
- Unit-Tests: Added missing unit-test for Stargazer Stellar Wallet XLM

* changes v6.2.0 -> v6.2.1

##
## Technical
##

- Dependencies: Updated unrar source from 5.9.4 to 6.0.5
- Dependencies: Make unrar dependencies optional and disable hash-mode 23800 if dependency is disabled

* changes v6.1.1 -> v6.2.0

##
## Algorithms
##

- Added hash-mode: Apple iWork
- Added hash-mode: AxCrypt 2 AES-128
- Added hash-mode: AxCrypt 2 AES-256
- Added hash-mode: BestCrypt v3 Volume Encryption
- Added hash-mode: Bitwarden
- Added hash-mode: Dahua Authentication MD5
- Added hash-mode: KNX IP Secure - Device Authentication Code
- Added hash-mode: MongoDB ServerKey SCRAM-SHA-1
- Added hash-mode: MongoDB ServerKey SCRAM-SHA-256
- Added hash-mode: Mozilla key3.db
- Added hash-mode: Mozilla key4.db
- Added hash-mode: MS Office 2016 - SheetProtection
- Added hash-mode: PDF 1.4 - 1.6 (Acrobat 5 - 8) - edit password
- Added hash-mode: PKCS#8 Private Keys
- Added hash-mode: RAR3-p (Compressed)
- Added hash-mode: RAR3-p (Uncompressed)
- Added hash-mode: RSA/DSA/EC/OPENSSH Private Keys
- Added hash-mode: SolarWinds Orion v2
- Added hash-mode: SolarWinds Serv-U
- Added hash-mode: SQLCipher
- Added hash-mode: Stargazer Stellar Wallet XLM
- Added hash-mode: Stuffit5
- Added hash-mode: Telegram Desktop >= v2.1.14 (PBKDF2-HMAC-SHA512)
- Added hash-mode: Umbraco HMAC-SHA1
- Added hash-mode: sha1($salt.sha1($pass.$salt))
- Added hash-mode: sha1(sha1($pass).$salt)

##
## Features
##

- Added new attack-mode: Association Attack (aka "Context Attack") to attack hashes from a hashlist with associated "hints"
- Added support for true UTF-8 to UTF-16 conversion in kernel crypto library
- Added option --hash-info to show generic information for each hash-mode
- Added command prompt [f]inish to tell hashcat to quit after finishing the current attack

##
## Bugs
##

- Fixed access to filename which is a null-pointer in benchmark mode
- Fixed both false negative and false positive results in -m 3000 in -a 3 (affecting only NVIDIA GPU)
- Fixed buffer overflow in -m 1800 in -O mode which is optimized to handle only password candidates up to length 15
- Fixed buffer overflow in -m 4710 in -P mode and only in single hash mode if salt length is larger than 32 bytes
- Fixed hardware management sysfs readings in status screen (typically ROCm controlled GPUs)
- Fixed include guards in several header files
- Fixed incorrect maximum password length support for -m 400 in optimized mode (reduced from 55 to 39)
- Fixed internal access on module option attribute OPTS_TYPE_SUGGEST_KG with the result that it was unused
- Fixed invalid handling of outfile folder entries for -m 22000
- Fixed memory leak causing problems in sessions with many iterations - for instance, --benchmark-all or large mask files
- Fixed memory leaks in several cases of errors with access to temporary files
- Fixed NVML initialization in WSL2 environments
- Fixed out-of-boundary reads in cases where user activates -S for fast but pure hashes in -a 1 or -a 3 mode
- Fixed out-of-boundary reads in kernels using module_extra_buffer_size() if -n is set to 1
- Fixed password reassembling for cracked hashes on host for slow hashes in optimized mode that are longer than 32 characters
- Fixed race condition in potfile check during removal of empty hashes
- Fixed race condition resulting in out of memory error on startup if multiple hashcat instances are started at the same time
- Fixed rare case of misalignment of the status prompt when other user warnings are shown in the hashcat output
- Fixed search of tuning database - if a device was not assigned an alias, it couldn't be found in general
- Fixed test on gzip header in wordlists and hashlists
- Fixed too-early execution of some module functions that use non-final values opts_type and opti_type
- Fixed unexpected non-unique salts in multi-hash cracking in Bitcoin/Litecoin wallet.dat module which led to false negatives
- Fixed unit test for -m 3000 by preventing it to generate zero hashes
- Fixed unit tests using 'null' as padding method in Crypt::CBC but actually want to use 'none'
- Fixed unterminated salt buffer in -m 23400 module_hash_encode() in case salt was of length 256
- Fixed vector datatype support in -m 21100 only -P mode and only -a 3 mode were affected

##
## Improvements
##

- Apple Keychain: Notify the user about the risk of collisions / false positives
- CUDA Backend: Do not warn about missing CUDA SDK installation if --backend-ignore-cuda is used
- CUDA Backend: Give detailed warning if either the NVIDIA CUDA or the NVIDIA RTC library cannot be initialized
- CUDA Backend: Use blocking events to avoid 100% CPU core usage (per GPU)
- OpenCL Runtime: Workaround JiT compiler deadlock on NVIDIA driver >= 465.89
- OpenCL Runtime: Workaround JiT compiler segfault on legacy AMDGPU driver compiling RAR3 OpenCL kernel
- RAR3 Kernels: Improved loop code, improving performance by 23%
- Scrypt Kernels: Added a number of GPU specific optimizations per hash modes to hashcat.hctune
- Scrypt Kernels: Added detailed documentation on device specific tunings in hashcat.hctune
- Scrypt Kernels: Optimized Salsa code portion by reducing register copies and removed unnecessary byte swaps
- Scrypt Kernels: Reduced kernel wait times by making it a true split kernel where iteration count = N value
- Scrypt Kernels: Refactored workload configuration strategy based on available resources
- Startup time: Improved startup time by avoiding some time-intensive operations for skipped devices

##
## Technical
##

- Bcrypt: Make BCRYPT entry for CPU in hashcat.hctune after switch to OPTS_TYPE_MP_MULTI_DISABLE (basically set -n to 1)
- Benchmark: Update benchmark_deep.pl with new hash modes added (also new hash modes which were added with v6.1.0)
- Building: Declare phony targets in Makefile to avoid conflicts of a target name with a file of the same name
- Building: Fixed build warnings on macOS for unrar sources
- Building: Fixed test for DARWIN_VERSION in Makefile
- Commandline Options: Removed option --example-hashes, now an alias of --hash-info
- Compute API: Skipping devices instead of stop if error occurred in initialization
- Documentation: Added 3rd party licenses to docs/license_libs
- Hash-Mode 8900 (Scrypt): Changed default benchmark scrypt parameters from 1k:1:1 to 16k:8:1 (default)
- Hash-Mode 11600 (7-Zip): Improved memory handling (alloc and free) for the hook function
- Hash-Mode 13200 (AxCrypt): Changed the name to AxCrypt 1 to avoid confusion
- Hash-Mode 13300 (AxCrypt in-memory SHA1): Changed the name to AxCrypt 1 in-memory SHA1
- Hash-Mode 16300 (Ethereum Pre-Sale Wallet, PBKDF2-HMAC-SHA256): Use correct buffer size allocation for AES key
- Hash-Mode 20710 (sha256(sha256($pass).$salt)): Removed unused code and fixed module_constraints
- Hash-Mode 22000 (WPA-PBKDF2-PMKID+EAPOL): Support loading a hash from command line
- Hash-Mode 23300 (Apple iWork): Use correct buffer size allocation for AES key
- Hash Parser: Output support for machine-readable hash lines in --show and --left and in error messages
- Kernel Development: Kernel cache is disabled automatically when hashcat is compiled with DEBUG=1
- Kernel Functions: Added generic AES-GCM interface see OpenCL/inc_cipher_aes-gcm.h
- Kernel Functions: Refactored OpenCL/inc_ecc_secp256k1.cl many functions, add constants and documentation
- Kernel Functions: Refactored OpenCL/inc_ecc_secp256k1.cl to improve usage in external programs
- Kernel Functions: Wrap atomic functions with hc_ prefix. Custom kernels need to rename "atomic_inc()" to "hc_atomic_inc()"
- Kernel Parameters: Added new parameter 'salt_repeat' to improve large buffer management
- Module Parameters: Add OPTS_TYPE_MP_MULTI_DISABLE for use by plugin developers to prevent multiply -n with the MCU count
- Module Parameters: Add OPTS_TYPE_NATIVE_THREADS for use by plugin developers to enforce native thread count
- Module Structure: Add 3rd party library hook management functions. This also requires an update to all existing module_init()
- OpenCL Runtime: Add support for clUnloadPlatformCompiler() to release some resources after JiT compilation
- OpenCL Runtime: Switched default OpenCL device type on macOS from GPU to CPU. Use -D 2 to enable GPU devices
- OpenCL Runtime: Update module_unstable_warnings() for all hash modes based on most recent versions of many OpenCL runtimes
- Unit tests: Added 'potthrough' (like passthrough, but hash:plain) to tools/test.pl
- Unit tests: Added Python 3 support for all of the Python code in our test framework
- Unit tests: Fixed the packaging of test (-p) feature
- Unit tests: Updated test.sh to show kernel type (pure or optimized) in output
- Unit tests: Use python3/pip3 instead of just python/pip in tools/install_modules.sh

* changes v6.1.0 -> v6.1.1

##
## Bugs
##

- Fixed unresolvable relative paths in hashcat.log

* changes v6.0.0 -> v6.1.0

##
## Algorithms
##

- Added hash-mode: Apple Keychain
- Added hash-mode: XMPP SCRAM

##
## Bugs
##

- Fixed alias detection with additional processor core count check
- Fixed false negatives in hash-mode 10901 if hash-mode 9200, 10000, 10900 or 20300 was used to compile the kernel binary
- Fixed integer overflow for large masks in -a 6 attack mode
- Fixed maximum password length in modules of hash-modes 600, 7800, 7801 and 9900
- Fixed non-zero status code when using --stdout
- Fixed uninitialized value in bitsliced DES kernel (BF mode only) leading to false negatives

##
## Improvements
##

- Compile macOS: Fixed makefile target 'clean' to correctly remove *.dSYM folders
- Compile ZLIB: Fixed makefile include paths in case USE_SYSTEM_ZLIB is used
- Hcchr Files: Renamed some .charset files into .hcchr files
- Hash-Mode 21200 (md5(sha1($salt).md5($pass))): Improved speed by using pre-computed SHA1
- OpenCL Devices: Utilize PCI domain to improve alias device detection
- OpenCL Kernels: Added datatypes to literals of enum constants
- OpenCL Kernels: Added pure kernels for hash-mode 600 (BLAKE2b-512)
- OpenCL Runtime: Add some unstable warnings for some SHA512 based algorithms on AMD GPU on macOS
- OpenCL Runtime: Reinterpret return code CL_DEVICE_NOT_FOUND from clGetDeviceIDs() as non-fatal

##
## Technical
##

- Backend: Changed the maximum number of compute devices from 64 to 128
- Tests: Improved tests for hash-mode 11300 (Bitcoin/Litecoin wallet.dat)
- Tests: Improved tests for hash-mode 13200 (AxCrypt)
- Tests: Improved tests for hash-mode 13600 (WinZip)
- Tests: Improved tests for hash-mode 16400 (CRAM-MD5 Dovecot)
- Tests: Improved tests for hash-mode 16800 (WPA-PMKID-PBKDF2)

* changes v5.1.0 -> v6.0.0

##
## Features
##

- Refactored hash-mode integration and replaced it with a fully modularized plugin interface
- Converted all existing hardwired hash-modes to hashcat plugins
- Added comprehensive plugin developer guide on adding new/custom hash-modes to hashcat
- Refactored compute backend interface to allow adding compute API other than OpenCL
- Added CUDA as a new compute backend (enables hashcat to run on NVIDIA Jetson, IBM POWER9 w/ Nvidia V100, etc.)
- Support automatic use of all available GPU memory when using CUDA backend
- Support automatic use of all available CPU cores for hash-mode-specific hooks
- Support on-the-fly loading of compressed wordlists in zip and gzip format
- Support deflate decompression for the 7-Zip hash-mode using zlib hook
- Added additional documentation on hashcat brain, slow-candidate interface and keyboard-layout mapping features
- Keep output of --show and --left in the original ordering of the input hash file
- Improved performance of many hash-modes

##
## Algorithms
##

- Added hash-mode: AES Crypt (SHA256)
- Added hash-mode: Android Backup
- Added hash-mode: AuthMe sha256
- Added hash-mode: BitLocker
- Added hash-mode: BitShares v0.x
- Added hash-mode: Blockchain, My Wallet, Second Password (SHA256)
- Added hash-mode: Citrix NetScaler (SHA512)
- Added hash-mode: Citrix NetScaler (PBKDF2-HMAC-SHA256)
- Added hash-mode: DiskCryptor
- Added hash-mode: Electrum Wallet (Salt-Type 3-5)
- Added hash-mode: Huawei Router sha1(md5($pass).$salt)
- Added hash-mode: Java Object hashCode()
- Added hash-mode: Kerberos 5 Pre-Auth etype 17 (AES128-CTS-HMAC-SHA1-96)
- Added hash-mode: Kerberos 5 Pre-Auth etype 18 (AES256-CTS-HMAC-SHA1-96)
- Added hash-mode: Kerberos 5 TGS-REP etype 17 (AES128-CTS-HMAC-SHA1-96)
- Added hash-mode: Kerberos 5 TGS-REP etype 18 (AES256-CTS-HMAC-SHA1-96)
- Added hash-mode: MultiBit Classic .key (MD5)
- Added hash-mode: MultiBit HD (scrypt)
- Added hash-mode: MySQL $A$ (sha256crypt)
- Added hash-mode: Open Document Format (ODF) 1.1 (SHA-1, Blowfish)
- Added hash-mode: Open Document Format (ODF) 1.2 (SHA-256, AES)
- Added hash-mode: Oracle Transportation Management (SHA256)
- Added hash-mode: PKZIP archive encryption
- Added hash-mode: PKZIP Master Key
- Added hash-mode: Python passlib pbkdf2-sha1
- Added hash-mode: Python passlib pbkdf2-sha256
- Added hash-mode: Python passlib pbkdf2-sha512
- Added hash-mode: QNX /etc/shadow (MD5)
- Added hash-mode: QNX /etc/shadow (SHA256)
- Added hash-mode: QNX /etc/shadow (SHA512)
- Added hash-mode: RedHat 389-DS LDAP (PBKDF2-HMAC-SHA256)
- Added hash-mode: Ruby on Rails Restful-Authentication
- Added hash-mode: SecureZIP AES-128
- Added hash-mode: SecureZIP AES-192
- Added hash-mode: SecureZIP AES-256
- Added hash-mode: SolarWinds Orion
- Added hash-mode: Telegram Desktop App Passcode (PBKDF2-HMAC-SHA1)
- Added hash-mode: Telegram Mobile App Passcode (SHA256)
- Added hash-mode: Web2py pbkdf2-sha512
- Added hash-mode: WPA-PBKDF2-PMKID+EAPOL
- Added hash-mode: WPA-PMK-PMKID+EAPOL
- Added hash-mode: md5($salt.sha1($salt.$pass))
- Added hash-mode: md5(sha1($pass).md5($pass).sha1($pass))
- Added hash-mode: md5(sha1($salt).md5($pass))
- Added hash-mode: sha1(md5(md5($pass)))
- Added hash-mode: sha1(md5($pass.$salt))
- Added hash-mode: sha1(md5($pass).$salt)
- Added hash-mode: sha1($salt1.$pass.$salt2)
- Added hash-mode: sha256(md5($pass))
- Added hash-mode: sha256($salt.$pass.$salt)
- Added hash-mode: sha256(sha256_bin($pass))
- Added hash-mode: sha256(sha256($pass).$salt)

##
## Bugs
##

- Fixed buffer overflow in build_plain() function
- Fixed buffer overflow in mp_add_cs_buf() function
- Fixed calculation of brain-session ID - only the first hash of the hashset was taken into account
- Fixed cleanup of password candidate buffers on GPU as set from autotune when -n parameter was used
- Fixed copy/paste error leading to invalid "Integer overflow detected in keyspace of mask" in attack-mode 6 and 7
- Fixed cracking multiple Office hashes (modes 9500, 9600) if hashes shared the same salt
- Fixed cracking of Blockchain, My Wallet (V1 and V2) hashes when testing decrypted data in unexpected format
- Fixed cracking of Cisco-PIX and Cisco-ASA MD5 passwords in mask-attack mode when mask > length 16
- Fixed cracking of DNSSEC (NSEC3) hashes by replacing all dots in the passwords with lengths
- Fixed cracking of Electrum Wallet Salt-Type 2 hashes
- Fixed cracking of NetNTLMv1 passwords in mask-attack mode when mask > length 16 (optimized kernels only)
- Fixed cracking of RAR3-hp hashes with pure kernel for passwords longer than 28 bytes
- Fixed cracking of VeraCrypt Streebog-512 hashes (CPU only)
- Fixed cracking raw Streebog-HMAC 256 and 512 hashes for passwords of length >= 64
- Fixed cracking raw Whirlpool hashes cracking for passwords of length >= 32
- Fixed incorrect progress-only result in a special race condition
- Fixed invalid call of mp_css_utf16le_expand()/mp_css_utf16be_expand() in slow-candidate sessions
- Fixed invalid password truncation in attack-mode 1 when the final password is longer than 32 characters
- Fixed invalid use of --hex-wordlist if encoded wordlist string is larger than length 256
- Fixed maximum password length limit which was announced as 256 but was actually 255
- Fixed out-of-boundary read in pure kernel rule engine rule 'p' when parameter was set to 2 or higher
- Fixed out-of-boundary write to decrypted[] in DPAPI masterkey file v1 kernel
- Fixed output of IKE PSK (mode 5300 and 5400) hashes to use separators in the correct position
- Fixed output password of "e" rule in pure and CPU rule engine when separator character is also the first letter
- Fixed problem with usage of hexadecimal notation (\x00-\xff) within rules
- Fixed race condition in maskfile mode by using a dedicated flag for restore execution
- Fixed some memory leaks when hashcat is shutting down due to some file error
- Fixed some memory leaks when mask-files are used in optimized mode
- Fixed --status-json to correctly escape certain characters in hashes
- Fixed the 7-Zip parser to allow the entire supported range of encrypted and decrypted data lengths
- Fixed the validation of the --brain-client-features command line argument (only values 1, 2 or 3 are allowed)

##
## Improvements
##

- Bitcoin Wallet: Be more user friendly by allowing a larger data range for ckey and public_key
- Brain: Added new parameter --brain-server-timer to specify seconds between scheduled backups
- Building: Fix for library compilation failure due to multiple definition of sbob_xx64()
- Cracking bcrypt and Password Safe v2: Use feedback from the compute API backend to dynamically calculate optimal thread count
- Dictstat: On Windows, the st_ino attribute in the stat struct is not set, which can lead to invalid cache hits. Added the filename to the database entry.
- Documents: Added README on how to build hashcat on Cygwin, MSYS2 and WSL
- File handling: Print a truncation warning when an oversized line is detected
- My Wallet: Added additional plaintext pattern used in newer versions
- Office cracking: Support hash format with second block data for 40-bit oldoffice files (eliminates false positives)
- OpenCL Runtime: Added a warning if OpenCL runtime NEO, Beignet, POCL (v1.4 or older) or MESA is detected, and skip associated devices (override with --force)
- OpenCL Runtime: Allow the kernel to access post-48k shared memory region on CUDA. Requires both module and kernel preparation
- OpenCL Runtime: Disable OpenCL kernel cache on Apple for Intel CPU (throws CL_BUILD_PROGRAM_FAILURE for no reason)
- OpenCL Runtime: Do not run shared- or constant-memory size checks if their memory type is of type global memory (typically CPU)
- OpenCL Runtime: Improve ROCm detection and make sure to not confuse with recent AMDGPU drivers
- OpenCL Runtime: Not using amd_bytealign (amd_bitalign is fine) on AMDGPU driver drastically reduces JiT segfaults
- OpenCL Runtime: Unlocked maximum thread count for NVIDIA GPU
- OpenCL Runtime: Update unstable mode warnings for Apple and AMDGPU drivers
- OpenCL Runtime: Workaround JiT compiler error on AMDGPU driver compiling WPA-EAPOL-PBKDF2 OpenCL kernel
- OpenCL Runtime: Workaround JiT compiler error on ROCm 2.3 driver if the 'inline' keyword is used in function declaration
- OpenCL Runtime: Workaround memory allocation error on AMD driver on Windows leading to CL_MEM_OBJECT_ALLOCATION_FAILURE
- OpenCL Runtime: Removed some workarounds by calling chdir() to specific folders on startup
- Outfile: Added new system to specify the outfile format, the new --outfile-format now also supports timestamps
- Startup Checks: Improved the pidfile check: Do not just check for existing PID, but also check executable filename
- Startup Checks: Prevent the user from modifying options which are overwritten automatically in benchmark mode
- Startup Screen: Add extra warning when using --force
- Startup Screen: Add extra warning when using --keep-guessing
- Startup Screen: Provide an estimate of host memory required for the requested attack
- Status Screen: Added brain status for all compute devices
- Status Screen: Added remaining counts and changed recovered count logic
- Status Screen: Added --status-json flag for easier machine reading of hashcat status output
- Tab Completion: Allow using "make install" version of hashcat
- Tuning Database: Updated hashcat.hctune with new models and refreshed vector width values
- VeraCrypt: Added support for VeraCrypt PIM brute-force, replaced --veracrypt-pim with --veracrypt-pim-start and --veracrypt-pim-stop
- WipZip cracking: Added two byte early reject, resulting in higher cracking speed
- WPA/WPA2 cracking: In the potfile, replace password with PMK in order to detect already cracked networks across all WPA modes

##
## Technical
##

- Backend Interface: Added new options --backend-ignore-cuda and --backend-ignore-opencl to prevent CUDA and/or OpenCL API from being used
- Binary Distribution: Removed 32-bit binary executables
- Building: On macOS, switch from ar to /usr/bin/ar to improve building compatibility
- Building: Skipping Travis/Appveyor build for non-code changes
- Codebase: Cleanup of many unused rc_* variables
- Codebase: Fixed some printf() format arguments
- Codebase: Fixed some type casting to avoid truncLongCastAssignment warnings
- Codebase: Moved hc_* file functions from shared.c to filehandling.c
- Codebase: Ran through a bunch of clang-tidy checkers and updated code accordingly
- Codebase: Remove redundant calls to fclose()
- Dependencies: Updated LZMA-Headers from 18.05 to 19.00
- Dependencies: Updated OpenCL-Headers to latest version from GitHub master repository
- Hash-Mode 12500 (RAR3-hp): Allow cracking of passwords up to length 64
- Hash-mode 1460 (HMAC-SHA256 (key = $salt)): Allow up to 64 byte of salt
- Hash-Mode 1680x (WPA-PMKID) specific: Changed separator character from '*' to ':'
- Hash-Mode 8300 (DNSSEC (NSEC3)) specific: Allow empty salt
- Keep Guessing: No longer automatically activate --keep-guessing for modes 9720, 9820, 14900 and 18100
- Keep Guessing: No longer mark hashes as cracked/removed when in potfile
- Kernel Cache: Reactivate OpenCL runtime specific kernel caches
- Kernel Compile: Removed -cl-std= from all kernel build options since we're compatible to all OpenCL versions
- OpenCL Kernels: Fix OpenCL compiler warning on double precision constants
- OpenCL Kernels: Moved "gpu_decompress", "gpu_memset" and "gpu_atinit" into shared.cl in order to reduce compile time
- OpenCL Options: Removed --opencl-platforms filter in order to force backend device numbers to stay constant
- OpenCL Options: Set --spin-damp to 0 (disabled) by default. With the CUDA backend this workaround became deprecated
- Parsers: switched from strtok() to strtok_r() for thread safety
- Requirements: Add new requirement for NVIDIA GPU: CUDA Toolkit (9.0 or later)
- Requirements: Update runtime check for minimum NVIDIA driver version from 367.x to 440.64 or later
- Test Script: Switched from /bin/bash to generic /bin/sh and updated code accordingly

* changes v5.0.0 -> v5.1.0

##
## Features
##

- Added support for using --stdout in brain-client mode
- Added new option --stdin-timeout-abort, to set how long hashcat should wait for stdin input before aborting
- Added new option --kernel-threads to manually override the automatically-calculated number of threads
- Added new option --keyboard-layout-mapping to map users keyboard layout, required to crack TC/VC system boot volumes

##
## Algorithms
##

- Added pure kernels for hash-mode 11700 (Streebog-256)
- Added pure kernels for hash-mode 11800 (Streebog-512)
- Added hash-mode 11750 (HMAC-Streebog-256 (key = $pass), big-endian)
- Added hash-mode 11760 (HMAC-Streebog-256 (key = $salt), big-endian)
- Added hash-mode 11850 (HMAC-Streebog-512 (key = $pass), big-endian)
- Added hash-mode 11860 (HMAC-Streebog-512 (key = $salt), big-endian)
- Added hash-mode 13771 (VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 512 bit)
- Added hash-mode 13772 (VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1024 bit)
- Added hash-mode 13773 (VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1536 bit)
- Added hash-mode 18200 (Kerberos 5 AS-REP etype 23)
- Added hash-mode 18300 (Apple File System (APFS))
- Added Kuznyechik cipher and cascades support for VeraCrypt kernels
- Added Camellia cipher and cascades support for VeraCrypt kernels

##
## Improvements
##

- OpenCL Devices: Add support for up to 64 OpenCL devices per system
- OpenCL Platforms: Add support for up to 64 OpenCL platforms per system
- OpenCL Runtime: Use our own yielding technique for synchronizing rather than vendor specific
- Startup: Show OpenCL runtime initialization message (per device)
- xxHash: Added support for using the version provided by the OS/distribution

##
## Bugs
##

- Fixed automated calculation of brain-session when not using all hashes in the hashlist
- Fixed calculation of brain-attack if a given wordlist has zero size
- Fixed checking the length of the last token in a hash if it was given the attribute TOKEN_ATTR_FIXED_LENGTH
- Fixed endianness and invalid separator character in outfile format for hash-mode 16801 (WPA-PMKID-PMK)
- Fixed ignoring --brain-client-features configuration when brain server has attack-position information from a previous run
- Fixed invalid hardware monitor detection in benchmark mode
- Fixed invalid warnings about throttling when --hwmon-disable was used
- Fixed missing call to WSACleanup() to cleanly shutdown windows sockets system
- Fixed missing call to WSAStartup() and client indexing in order to start the brain server on Windows
- Fixed out-of-boundary read in DPAPI masterkey file v2 OpenCL kernel
- Fixed out-of-bounds write in short-term memory of the brain server
- Fixed output of --speed-only and --progress-only when fast hashes are used in combination with --slow-candidates
- Fixed selection of OpenCL devices (-d) if there's more than 32 OpenCL devices installed
- Fixed status output of progress value when -S and -l are used in combination
- Fixed thread count maximum for pure kernels in straight attack mode

##
## Technical
##

- Brain: Set --brain-client-features default from 3 to 2
- Dependencies: Added xxHash and OpenCL-Headers to deps/ in order to allow building hashcat from GitHub source release package
- Dependencies: Removed gitmodules xxHash and OpenCL-Headers
- Keymaps: Added hashcat keyboard mapping us.hckmap (can be used as template)
- Keymaps: Added hashcat keyboard mapping de.hckmap
- Hardware Monitor: Renamed --gpu-temp-abort to --hwmon-temp-abort
- Hardware Monitor: Renamed --gpu-temp-disable to --hwmon-disable
- Memory: Limit maximum host memory allocation depending on bitness
- Memory: Reduced default maximum bitmap size from 24 to 18 and give a notice to use --bitmap-max to restore
- Parameter: Rename --nvidia-spin-damp to --spin-damp (now accessible for all devices)
- Pidfile: Treat a corrupted pidfile like a not existing pidfile
- OpenCL Device: Do a real query on OpenCL local memory type instead of just assuming it
- OpenCL Runtime: Disable auto-vectorization for Intel OpenCL runtime to workaround hanging JiT since version 18.1.0.013
- Tests: Added hash-mode 11700 (Streebog-256)
- Tests: Added hash-mode 11750 (HMAC-Streebog-256 (key = $pass), big-endian)
- Tests: Added hash-mode 11760 (HMAC-Streebog-256 (key = $salt), big-endian)
- Tests: Added hash-mode 11800 (Streebog-512)
- Tests: Added hash-mode 11850 (HMAC-Streebog-512 (key = $pass), big-endian)
- Tests: Added hash-mode 11860 (HMAC-Streebog-512 (key = $salt), big-endian)
- Tests: Added hash-mode 13711 (VeraCrypt PBKDF2-HMAC-RIPEMD160 + XTS 512 bit)
- Tests: Added hash-mode 13712 (VeraCrypt PBKDF2-HMAC-RIPEMD160 + XTS 1024 bit)
- Tests: Added hash-mode 13713 (VeraCrypt PBKDF2-HMAC-RIPEMD160 + XTS 1536 bit)
- Tests: Added hash-mode 13721 (VeraCrypt PBKDF2-HMAC-SHA512 + XTS 512 bit)
- Tests: Added hash-mode 13722 (VeraCrypt PBKDF2-HMAC-SHA512 + XTS 1024 bit)
- Tests: Added hash-mode 13723 (VeraCrypt PBKDF2-HMAC-SHA512 + XTS 1536 bit)
- Tests: Added hash-mode 13731 (VeraCrypt PBKDF2-HMAC-Whirlpool + XTS 512 bit)
- Tests: Added hash-mode 13732 (VeraCrypt PBKDF2-HMAC-Whirlpool + XTS 1024 bit)
- Tests: Added hash-mode 13733 (VeraCrypt PBKDF2-HMAC-Whirlpool + XTS 1536 bit)
- Tests: Added hash-mode 13751 (VeraCrypt PBKDF2-HMAC-SHA256 + XTS 512 bit)
- Tests: Added hash-mode 13752 (VeraCrypt PBKDF2-HMAC-SHA256 + XTS 1024 bit)
- Tests: Added hash-mode 13753 (VeraCrypt PBKDF2-HMAC-SHA256 + XTS 1536 bit)
- Tests: Added hash-mode 13771 (VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 512 bit)
- Tests: Added hash-mode 13772 (VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1024 bit)
- Tests: Added hash-mode 13773 (VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1536 bit)
- Tests: Added VeraCrypt containers for Kuznyechik cipher and cascades
- Tests: Added VeraCrypt containers for Camellia cipher and cascades

* changes v4.2.1 -> v5.0.0

##
## Features
##

- Added new option --slow-candidates which allows hashcat to generate passwords on-host
- Added new option --brain-server to start a hashcat brain server
- Added new option --brain-client to start a hashcat brain client, automatically activates --slow-candidates
- Added new option --brain-host and --brain-port to specify ip and port of brain server, both listening and connecting
- Added new option --brain-session to override automatically calculated brain session ID
- Added new option --brain-session-whitelist to allow only explicit written session ID on brain server
- Added new option --brain-password to specify the brain server authentication password
- Added new option --brain-client-features which allows enable and disable certain features of the hashcat brain

##
## Algorithms
##

- Added hash-mode 17300 = SHA3-224
- Added hash-mode 17400 = SHA3-256
- Added hash-mode 17500 = SHA3-384
- Added hash-mode 17600 = SHA3-512
- Added hash-mode 17700 = Keccak-224
- Added hash-mode 17800 = Keccak-256
- Added hash-mode 17900 = Keccak-384
- Added hash-mode 18000 = Keccak-512
- Added hash-mode 18100 = TOTP (HMAC-SHA1)
- Removed hash-mode 5000 = SHA-3 (Keccak)

##
## Improvements
##

- Added additional hybrid "passthrough" rules, to enable variable-length append/prepend attacks
- Added a periodic check for read timeouts in stdin/pipe mode, and abort if no input was provided
- Added a tracker for salts, amplifier and iterations to the status screen
- Added option --markov-hcstat2 to make it clear that the new hcstat2 format (compressed hcstat2gen output) must be used
- Allow bitcoin master key lengths other than 96 bytes (but they must be always multiples of 16)
- Allow hashfile for -m 16800 to be used with -m 16801
- Allow keepass iteration count to be larger than 999999
- Changed algorithms using colon as separators in the hash to not use the hashconfig separator on parsing
- Do not allocate memory segments for bitmap tables if we don't need it - for example, in benchmark mode
- Got rid of OPTS_TYPE_HASH_COPY for Ansible Vault
- Improved the speed of the outfile folder scan when using many hashes/salts
- Increased the maximum size of edata2 in Kerberos 5 TGS-REP etype 23
- Make the masks parser more restrictive by rejecting a single '?' at the end of the mask (use ?? instead)
- Override --quiet and show final status screen in case --status is used
- Removed duplicate words in the dictionary file example.dict
- Updated Intel OpenCL runtime version check
- Work around some AMD OpenCL runtime segmentation faults
- Work around some padding issues with host compilers and OpenCL JiT on 32 and 64-bit systems

##
## Bugs
##

- Fixed a invalid scalar datatype return value in hc_bytealign() where it should be a vector datatype return value
- Fixed a problem with attack mode -a 7 together with stdout mode where the mask bytes were missing in the output
- Fixed a problem with tab completion where --self-test-disable incorrectly expected a further parameter/value
- Fixed a race condition in status view that lead to out-of-bound reads
- Fixed detection of unique ESSID in WPA-PMKID-* parser
- Fixed missing wordlist encoding in combinator mode
- Fixed speed/delay problem when quitting while the outfile folder is being scanned
- Fixed the ciphertext max length in Ansible Vault parser
- Fixed the tokenizer configuration in Postgres hash parser
- Fixed the byte order of digest output for hash-mode 11800 (Streebog-512)

* changes v4.2.0 -> v4.2.1

##
## Improvements
##

- Try to evaluate available OpenCL device memory and use this information instead of total available OpenCL device memory for autotune

##
## Bugs
##

- Fixed a buffer overflow in precompute_salt_md5() in case salt was longer than 64 characters

* changes v4.1.0 -> v4.2.0

##
## Algorithms
##

- Added hash-mode 16700 = FileVault 2
- Added hash-mode 16800 = WPA-PMKID-PBKDF2
- Added hash-mode 16801 = WPA-PMKID-PMK
- Added hash-mode 16900 = Ansible Vault

##
## Improvements
##

- Added JtR-compatible support for hex notation in rules engine
- Added OpenCL device utilization to the status information in machine-readable output
- Added missing NV Tesla and Titan GPU details to tuning database
- General file handling: Abort if a byte order mark (BOM) was detected in a wordlist, hashlist, maskfile or rulefile
- HCCAPX management: Use advanced hints in message_pair stored by hcxtools about endian bitness of replay counter
- OpenCL kernels: Abort session if kernel self-test fails
- OpenCL kernels: Add '-pure' prefix to kernel filenames to avoid problems caused by reusing existing hashcat installation folder
- OpenCL kernels: Removed the use of 'volatile' in inline assembly instructions where it is not needed
- OpenCL kernels: Switched array pointer types in function declarations in order to be compatible with OpenCL 2.0
- Refactored code for --progress-only and --speed-only calculation
- SIP cracking: Increased the nonce field to allow a salt of 1024 bytes
- TrueCrypt/VeraCrypt cracking: Do an entropy check on the TC/VC header on start

##
## Bugs
##

- Fixed a function declaration attribute in -m 8900 kernel leading to unusable -m 9300 which shares kernel code with -m 8900
- Fixed a miscalculation in --progress-only mode output for extremely slow kernels like -m 14800
- Fixed a missing check for errors on OpenCL devices leading to invalid removal of restore file
- Fixed a missing kernel in -m 5600 in combination with -a 3 and -O if mask is >= 16 characters
- Fixed detection of AMD_GCN version in case the rocm driver is used
- Fixed missing code section in -m 2500 and -m 2501 to crack corrupted handshakes with a LE endian bitness base
- Fixed a missing check for hashmodes using OPTS_TYPE_PT_UPPER causing the self-test to fail when using combinator- and hybrid-mode

* changes v4.0.1 -> v4.1.0

##
## Features
##

- Added option --benchmark-all to benchmark all hash-modes (not just the default selection)
- Removed option --gpu-temp-retain that tried to retain GPU temperature at X degrees celsius - please use driver-specific tools
- Removed option --powertune-enable to enable power tuning - please use driver specific tools

##
## Algorithms
##

- Added hash-mode 16000 = Tripcode
- Added hash-mode 16100 = TACACS+
- Added hash-mode 16200 = Apple Secure Notes
- Added hash-mode 16300 = Ethereum Pre-Sale Wallet, PBKDF2-SHA256
- Added hash-mode 16400 = CRAM-MD5 Dovecot
- Added hash-mode 16500 = JWT (JSON Web Token)
- Added hash-mode 16600 = Electrum Wallet (Salt-Type 1-2)

##
## Bugs
##

- Fixed a configuration setting for -m 400 in pure kernel mode which said it was capable of doing SIMD when it is not
- Fixed a hash parsing problem for 7-Zip hashes: allow a longer CRC32 data length field within the hash format
- Fixed a hash parsing problem when using --show/--left with hashes with long salts that required pure kernels
- Fixed a logic error in storing temporary progress for slow hashes, leading to invalid speeds in status view
- Fixed a mask-length check issue: return -1 in case the mask length is not within the password-length range
- Fixed a missing check for return code in case hashcat.hcstat2 was not found
- Fixed a race condition in combinator- and hybrid-mode where the same scratch buffer was used by multiple threads
- Fixed a restore issue leading to "Restore value is greater than keyspace" when mask files or wordlist folders were used
- Fixed a uninitialized value in OpenCL kernels 9720, 9820 and 10420 leading to absurd benchmark performance
- Fixed the maximum password length check in password-reassembling function
- Fixed the output of --show when $HEX[] passwords were present within the potfile

##
## Improvements
##

- OpenCL Kernels: Add a decompressing kernel and a compressing host code in order to reduce PCIe transfer time
- OpenCL Kernels: Improve performance preview accuracy in --benchmark, --speed-only and --progress-only mode
- OpenCL Kernels: Remove password length restriction of 16 for Cisco-PIX and Cisco-ASA hashes
- Terminal: Display set cost/rounds during benchmarking
- Terminal: Show [r]esume in prompt only in pause mode, and show [p]ause in prompt only in resume mode

##
## Technical
##

- Autotune: Improve autotune engine logic and synchronize results on same OpenCL devices
- Documents: Added docs/limits.txt
- Files: Copy include/ folder and its content when SHARED is set to 1 in Makefile
- Files: Switched back to relative current working directory on windows to work around problems with Unicode characters
- Hashcat Context: Fixed a memory leak in shutdown phase
- Hash Parser: Changed the way large strings are handled/truncated within the event buffer if they are too large to fit
- Hash Parser: Fixed a memory leak in shutdown phase
- Hash Parser: Fixed the use of strtok_r () calls
- OpenCL Devices: Fixed several memory leaks in shutdown phase
- OpenCL Kernels: Add general function declaration keyword (inline) and some OpenCL runtime specific exceptions for NV and CPU devices
- OpenCL Kernels: Replace variables from uXX to uXXa if used in __constant space
- OpenCL Kernels: Use a special kernel to initialize the password buffer used during autotune measurements, to reduce startup time
- OpenCL Kernels: Refactored kernel thread management from native to maximum per kernel
- OpenCL Kernels: Use three separate comparison kernels (depending on keyver) for WPA instead of one
- OpenCL Runtime: Add current timestamp to OpenCL kernel source in order to force OpenCL JiT compiler to recompile and not use the cache
- OpenCL Runtime: Enforce use of OpenCL version 1.2 to restrain OpenCL runtimes to make use of the __generic address space qualifier
- OpenCL Runtime: Updated rocm detection
- Returncode: Enforce return code 0 when the user selects --speed-only or --progress-only and no other error occurs
- Rules: Fixed some default rule-files after changing rule meaning of 'x' to 'O'
- Self Test: Skip self-test for mode 8900 - user-configurable scrypt settings are incompatible with fixed settings in the self-test hash
- Self Test: Skip self-test for mode 15700 because the settings are too high and cause startup times that are too long
- Terminal: Add workitem settings to status display (can be handy for debugging)
- Terminal: Send clear-line code to the same output stream as the message immediately following
- Timer: Switch from gettimeofday() to clock_gettime() to work around problems on cygwin
- User Options: According to getopts manpage, the last element of the option array has to be filled with zeros

* changes v4.0.0 -> v4.0.1:

##
## Improvements
##

- Changed the maximum length of the substring of a hash shown whenever the parser found a problem while parsing the hash

##
## Bugs
##

- Fixed a memory leak while parsing a wordlist
- Fixed compile of kernels on AMD systems on windows due to invalid detection of ROCm
- Fixed compile of sources using clang under MSYS2
- Fixed overlapping memory segment copy in CPU rule engine if using a specific rule function
- Fixed a parallel build problem when using the "install" Makefile target
- Fixed the version number extraction for github releases which do not including the .git directory

* changes v3.6.0 -> v4.0.0:

##
## Features
##

- Added support to crack passwords and salts up to length 256
- Added option --optimized-kernel-enable to use faster kernels but limit the maximum supported password- and salt-length
- Added self-test functionality to detect broken OpenCL runtimes on startup
- Added option --self-test-disable to disable self-test functionality on startup
- Added option --wordlist-autohex-disable to disable the automatic conversion of $HEX[] words from the word list
- Added option --example-hashes to show an example hash for each hash-mode
- Removed option --weak-hash-check (zero-length password check) to increase startup time, it also causes many Trap 6 error on macOS

##
## Algorithms
##

- Added hash-mode 2500 = WPA/WPA2 (SHA256-AES-CMAC)
- Added hash-mode 2501 = WPA/WPA2 PMK

##
## Bugs
##

- Fixed a buffer overflow in mangle_dupechar_last function
- Fixed a calculation error in get_power() leading to errors of type "BUG pw_add()!!"
- Fixed a memory problem that occurred when the OpenCL folder was not found and e.g. the shared and session folder were the same
- Fixed a missing barrier() call in the RACF OpenCL kernel
- Fixed a missing salt length value in benchmark mode for SIP
- Fixed an integer overflow in hash buffer size calculation
- Fixed an integer overflow in innerloop_step and innerloop_cnt variables
- Fixed an integer overflow in masks not skipped when loaded from file
- Fixed an invalid optimization code in kernel 7700 depending on the input hash, causing the kernel to loop forever
- Fixed an invalid progress value in status view if words from the base wordlist get rejected because of length
- Fixed a parser error for mode -m 9820 = MS Office <= 2003 $3, SHA1 + RC4, collider #2
- Fixed a parser error in multiple modes not checking for return code, resulting in negative memory index writes
- Fixed a problem with changed current working directory, for instance by using --restore together with --remove
- Fixed a problem with the conversion to the $HEX[] format: convert/hexify also all passwords of the format $HEX[]
- Fixed the calculation of device_name_chksum; should be done for each iteration
- Fixed the dictstat lookup if nanoseconds are used in timestamps for the cached files
- Fixed the estimated time value whenever the value is very large and overflows
- Fixed the output of --show when used together with the collider modes -m 9710, 9810 or 10410
- Fixed the parsing of command line options. It doesn't show two times the same error about an invalid option anymore
- Fixed the parsing of DCC2 hashes by allowing the "#" character within the user name
- Fixed the parsing of descrypt hashes if the hashes do have non-standard characters within the salt
- Fixed the use of --veracrypt-pim option. It was completely ignored without showing an error
- Fixed the version number used in the restore file header

##
## Improvements
##

- Autotune: Do a pre-autotune test run to find out if kernel runtime is above some TDR limit
- Charset: Add additional DES charsets with corrected parity
- OpenCL Buffers: Do not allocate memory for amplifiers for fast hashes, it's simply not needed
- OpenCL Kernels: Improved performance of SHA-3 Kernel (keccak) by hardcoding the 0x80 stopbit
- OpenCL Kernels: Improved rule engine performance by 6% on for NVidia
- OpenCL Kernels: Move from ld.global.v4.u32 to ld.const.v4.u32 in _a3 kernels
- OpenCL Kernels: Replace bitwise swaps with rotate() versions for AMD
- OpenCL Kernels: Rewritten Keccak kernel to run fully on registers and partially reversed last round
- OpenCL Kernels: Rewritten SIP kernel from scratch
- OpenCL Kernels: Thread-count is set to hardware native count except if -w 4 is used then OpenCL maximum is used
- OpenCL Kernels: Updated default scrypt TMTO to be ideal for latest NVidia and AMD top models
- OpenCL Kernels: Vectorized tons of slow kernels to improve CPU cracking speed
- OpenCL Runtime: Improved detection for AMD and NV devices on macOS
- OpenCL Runtime: Improved performance on Intel MIC devices (Xeon PHI) on runtime level (300MH/s to 2000MH/s)
- OpenCL Runtime: Updated AMD ROCm driver version check, warn if version < 1.1
- Show cracks: Improved the performance of --show/--left if used together with --username
- Startup: Add visual indicator of active options when benchmarking
- Startup: Check and abort session if outfile and wordlist point to the same file
- Startup: Show some attack-specific optimizer constraints on start, eg: minimum and maximum support password- and salt-length
- WPA cracking: Improved nonce-error-corrections mode to use a both positive and negative corrections

##
## Technical
##

- General: Update C standard from c99 to gnu99
- Hash Parser: Improved salt-length checks for generic hash modes
- HCdict File: Renamed file from hashcat.hcdict to hashcat.hcdict2 and add header because versions are incompatible
- HCstat File: Add code to read LZMA compressed hashcat.hcstat2
- HCstat File: Add hcstat2 support to enable masks of length up to 256, also adds a filetype header
- HCstat File: Renamed file from hashcat.hcstat to hashcat.hcstat2 and add header because versions are incompatible
- HCtune File: Remove apple related GPU entries to workaround Trap 6 error
- OpenCL Kernels: Added code generator for most of the switch_* functions and replaced existing code
- OpenCL Kernels: Declared all include functions as static to reduce binary kernel cache size
- OpenCL Kernels: On AMD GPU, optimized kernels for use with AMD ROCm driver
- OpenCL Kernels: Removed some include functions that are no longer needed to reduce compile time
- OpenCL Runtime: Fall back to 64 threads default (from 256) on AMD GPU to prevent creating too many workitems
- OpenCL Runtime: Forcing OpenCL 1.2 no longer needed. Option removed from build options
- OpenCL Runtime: On AMD GPU, recommend AMD ROCm driver for Linux
- Restore: Fixed the version number used in the restore file header
- Time: added new type for time measurements hc_time_t and related functions to force the use of 64 bit times

* changes v3.5.0 -> v3.6.0:

##
## Algorithms
##

- Added hash-mode   600 = BLAKE2-512
- Added hash-mode 15200 = Blockchain, My Wallet, V2
- Added hash-mode 15300 = DPAPI masterkey file v1 and v2
- Added hash-mode 15400 = ChaCha20
- Added hash-mode 15500 = JKS Java Key Store Private Keys (SHA1)
- Added hash-mode 15600 = Ethereum Wallet, PBKDF2-HMAC-SHA256
- Added hash-mode 15700 = Ethereum Wallet, PBKDF2-SCRYPT

##
## Features
##

- 7-Zip cracking: increased max. data length to 320k and removed AES padding attack to avoid false negatives
- Dictionary cache: Show time spent on dictionary cache building at startup
- Rules: Support added for position 'p' (Nth instance of a character) in host mode (using -j or -k)
- Rules: Support added for rejection rule '_N' (reject plains of length not equal to N) in host mode
- Rules: Support added for rule 'eX'
- Wordlist encoding: Added parameters --encoding-from and --encoding-to to configure wordlist encoding handling
- Wordlist encoding: Support added for internal conversion between user-defined encodings during runtime

##
## Workarounds
##

- Workaround added for NVIDIA NVML library: If libnvidia-ml.so couldn't be loaded, try again using libnvidia-ml.so.1

##
## Improvements
##

- WPA cracking: Improved nonce-error-corrections mode to fix corrupt nonces generated on big-endian devices

##
## Bugs
##

- Fixed a condition that caused hybrid attacks using a maskfile to not select all wordlists from a wordlist folder
- Fixed a memory leak that was present when a user periodically prints hashcat status (using --status-timer)
- Fixed a missing type specifier in a function declaration of the RACF kernel

##
## Technical
##

- Building: In the binary release packages, link libiconv static for Windows binaries
- Dictstat: Structure for dictstat file changed to include --encoding-from and --encoding-to parameters
- OpenCL Runtime: Updated AMDGPU-PRO driver version check, warn if version 17.10 (known to be broken) is detected
- WPA cracking: Reduced --nonce-error-corrections default from 16 to 8 to compensate for speed drop caused by big-endian fixes

* changes v3.40 -> v3.5.0:

##
## Features
##

- WPA cracking: Added support for WPA/WPA2 handshake AP nonce automatic error correction
- WPA cracking: Added parameter --nonce-error-corrections to configure range of error correction

##
## Algorithms
##

- Added hash-mode 15100 = Juniper/NetBSD sha1crypt

##
## Improvements
##

- Abbreviate long hashes to display the Hash.Target status line within 80 characters
- Refactored internal use of esalt to sync with the number of digests instead of the number of salts
- Refactored other output to display within 80 characters without wrapping

##
## Bugs
##

- Fixed a hash validation error when trying to load Android FDE < 4.3 hashes
- Fixed a problem where --keyspace combined with custom charsets incorrectly displayed an error message
- Fixed a problem where --stdout combined with custom charsets incorrectly displayed an error message
- Fixed a problem with parsing and displaying -m 7000 = Fortigate (FortiOS) hashes
- Fixed a race condition after sessions finish, where the input-base was freed but accessed afterwards
- Fixed a typo that resulted in the minimum password length not being correctly initialized
- Fixed --outfile-format formats 11 through 15 to show the correct crack position
- Fixed --remove to apply even when all hashes are either found in the potfile or detected in weak-hash checks

##
## Technical
##

- Building: Added missing prototypes for atlassian_parse_hash function
- Dictionary Cache: Split long status line into multiple lines to stay < 80 chars
- Files: Detect and error when users try to use -r with a parameter which is not a file
- HCCAPX Parser: Added support for a special bit (bit 8) of the message_pair that indicates if replay counters match
- Parameter: Detect and error when users try to use an empty string (length 0) for parameters like --session=
- Parameter: Detect and error when users try to use non-digit input when only digits are expected
- Sessions: Improved string comparison in case user sets --session to "hashcat"
- Status View: Add rejected counter to machine-readable output
- Status View: Rename labels Input.Mode, Input.Base, ... to Guess.Mode, Guess.Base, ...
- Status View: Added a visual indicator to the status screen when checkpoint quit has been requested
- Versions: Changed version naming convention from x.yz to x.y.z

* changes v3.30 -> v3.40:

##
## Features
##

- Added support for loading hccapx files
- Added support for filtering hccapx message pairs using --hccapx-message-pair
- Added support for parsing 7-Zip hashes with LZMA/LZMA2 compression indicator set to a non-zero value
- Added support for decompressing LZMA1/LZMA2 data for -m 11600 = 7-Zip to validate the CRC
- Added support for automatic merge of LM halves in case --show and --left is used
- Added support for showing all user names with --show and --left if --username was specified
- Added support for GPU temperature management on cygwin build

##
## Algorithms
##

- Added hash-mode  1411 = SSHA-256(Base64), LDAP {SSHA256}
- Added hash-mode  3910 = md5(md5($pass).md5($salt))
- Added hash-mode  4010 = md5($salt.md5($salt.$pass))
- Added hash-mode  4110 = md5($salt.md5($pass.$salt))
- Added hash-mode  4520 = sha1($salt.sha1($pass))
- Added hash-mode  4522 = PunBB
- Added hash-mode  7000 = Fortigate (FortiOS)
- Added hash-mode 12001 = Atlassian (PBKDF2-HMAC-SHA1)
- Added hash-mode 14600 = LUKS
- Added hash-mode 14700 = iTunes Backup < 10.0
- Added hash-mode 14800 = iTunes Backup >= 10.0
- Added hash-mode 14900 = Skip32
- Added hash-mode 15000 = FileZilla Server >= 0.9.55

##
## Workarounds
##

- Workaround added for AMDGPU-Pro OpenCL runtime: AES encrypt and decrypt Invertkey function was calculated wrong in certain cases
- Workaround added for AMDGPU-Pro OpenCL runtime: RAR3 kernel require a volatile variable to work correctly
- Workaround added for Apple OpenCL runtime: bcrypt kernel requires a volatile variable because of a compiler optimization bug
- Workaround added for Apple OpenCL runtime: LUKS kernel requires some volatile variables because of a compiler optimization bug
- Workaround added for Apple OpenCL runtime: TrueCrypt kernel requires some volatile variables because of a compiler optimization bug
- Workaround added for NVidia OpenCL runtime: RACF kernel requires EBCDIC lookup to be done on shared memory

##
## Bugs
##

- Fixed a problem within the Kerberos 5 TGS-REP (-m 13100) hash parser
- Fixed clEnqueueNDRangeKernel(): CL_UNKNOWN_ERROR caused by an invalid work-item count during weak-hash-check
- Fixed cracking of PeopleSoft Token (-m 13500) if salt length + password length is >= 128 byte
- Fixed cracking of Plaintext (-m 99999) in case MD4 was used in a previous session
- Fixed DEScrypt cracking in BF mode in case the hashlist contains more than 16 times the same salt
- Fixed duplicate detection for WPA handshakes with the same ESSID
- Fixed nvapi datatype definition for NvS32 and NvU32
- Fixed overflow in bcrypt kernel in expand_key() function
- Fixed pointer to local variable outside scope in case -j or -k is used
- Fixed pointer to local variable outside scope in case --markov-hcstat is not used
- Fixed recursion in loopback handling when session was aborted by the user
- Fixed rule 'O' (RULE_OP_MANGLE_OMIT) in host mode in case the offset + length parameter equals the length of the input word
- Fixed rule 'i' (RULE_OP_MANGLE_INSERT) in host mode in case the offset parameter equals the length of the input word
- Fixed string not null terminated inside workaround for checking drm driver path
- Fixed string not null terminated while reading maskfiles
- Fixed truncation of password after position 32 with the combinator attack
- Fixed use of option --keyspace in combination with -m 2500 (WPA)
- Fixed WPA/WPA2 cracking in case eapol frame is >= 248 byte

##
## Technical
##

- Building: Add SHARED variable to Makefile to choose if hashcat is build as static or shared binary (using libhashcat.so/hashcat.dll)
- Building: Removed compiler option -march=native as this created problems for maintainers on various distributions
- Building: Removed the use of RPATH on linker level
- Building: Replaced linking of CRT_glob.o with the use of int _dowildcard
- Commandline: Do some checks related to custom-charset options if user specifies them
- CPU Affinity: Fixed memory leak in case invalid cpu Id was specified
- Dispatcher: Fixed several memory leaks in case an OpenCL error occurs
- Events: Improved the maximum event message handling. event_log () will now also internally make sure that the message is properly terminated
- File Locking: Improved error detection on file locks
- File Reads: Fixed memory leak in case outfile or hashfile was not accessible
- File Reads: Improved error detection on file reads, especially when getting the file stats
- Files: Do several file and folder checks on startup rather than when they are actually used to avoid related error after eventual intense operations
- Hardware Management: Bring back kernel exec timeout detection for NVidia on user request
- Hardware Monitor: Fixed several memory leaks in case hash-file writing (caused by --remove) failed
- Hardware Monitor: Fixed several memory leaks in case no hardware monitor sensor is found
- Hardware Monitor: In case NVML initialization failed, do not try to initialize NVAPI or XNVCTRL because they both depend on NVML
- Hash Parsing: Added additional bound checks for the SIP digest authentication (MD5) parser (-m 11400)
- Hash Parsing: Make sure that all files are correctly closed whenever a hash file parsing error occurs
- Helper: Added functions to check existence, type, read- and write-permissions and rewrite sources to use them instead of stat()
- Keyfile handling: Make sure that the memory is cleanly freed whenever a VeraCrypt/TrueCrypt keyfile fails to load
- Mask Checks: Added additional memory cleanups after parsing/verifying masks
- Mask Checks: Added integer overflow detection for a keyspace of a mask provided by user
- Mask Increment: Fixed memory leak in case mask_append() fails
- OpenCL Device: Do a check on available constant memory size and abort if it's less than 64kB
- OpenCL Device Management: Fixed several memory leaks in case initialization of an OpenCL device or platform failed
- OpenCL Header: Updated CL_* errorcode to OpenCL 1.2 standard
- OpenCL Kernel: Move kernel binary buffer from heap to stack memory
- OpenCL Kernel: Refactored read_kernel_binary to load only a single kernel for a single device
- OpenCL Kernel: Remove "static" keyword from function declarations; Causes older Intel OpenCL runtimes to fail compiling
- OpenCL Kernel: Renumbered hash-mode 7600 to 4521
- OpenCL Runtime: Added a warning about using Mesa OpenCL runtime
- OpenCL Runtime: Updated AMDGPU-Pro driver version check, do warn if version 16.60 is detected which is known to be broken
- Outfile Check: Fixed a memory leak for failed outfile reads
- Restore: Add some checks on the rd->cwd variable in restore case
- Rule Engine: Fixed several memory leaks in case loading of rules failed
- Session Management: Automatically set dedicated session names for non-cracking parameters, for example: --stdout
- Session Management: Fixed several memory leaks in case profile- or install-folder setup failed
- Sessions: Move out handling of multiple instance from restore file into separate pidfile
- Status screen: Do not try to clear prompt in --quiet mode
- Tests: Fixed the timeout status code value and increased the runtime to 400 seconds
- Threads: Restored strerror as %m is unsupported by the BSDs
- Wordlists: Disable dictstat handling for hash-mode 3000 as it virtually creates words in the wordlist which is not the case for other modes
- Wordlists: Fixed memory leak in case access a file in a wordlist folder fails
- WPA: Changed format for outfile and potfile from essid:mac1:mac2 to hash:mac_ap:mac_sta:essid
- WPA: Changed format for outfile_check from essid:mac1:mac2 to hash

* changes v3.20 -> v3.30:

##
## Features
##

- Files: Use $HEX[...] in case the password includes the separator character, increases potfile reading performance
- Files: If the user specifies a folder to scan for wordlists instead of directly a wordlist, then ignore the hidden files
- Loopback: Include passwords for removed hashes present in the potfile to next loopback iteration
- New option --progress-only: Quickly provides ideal progress step size and time to process on the user hashes and selected options, then quit
- Status screen: Reenabled automatic status screen display in case of stdin used
- Truecrypt/Veracrypt: Use CRC32 to verify headers instead of fuzzy logic, greatly reduces false positives from 18:2^48 to 3:2^64
- WPA cracking: Reuse PBKDF2 intermediate keys if duplicate essid is detected

##
## Algorithms
##

- Added hash-mode 1300 = SHA-224

##
## Bugs
##

- Fixed buffer overflow in status screen display in case of long non-utf8 string
- Fixed buffer overflow in plaintext parsing code: Leading to segfault
- Fixed custom char parsing code in maskfiles in --increment mode: Custom charset wasn't used
- Fixed display screen to show input queue when using custom charset or rules
- Fixed double fclose() using AMDGPU-Pro on sysfs compatible platform: Leading to segfault
- Fixed hash-mode 11400 = SIP digest authentication (MD5): Cracking of hashes which did not include *auth* or *auth-int* was broken
- Fixed hex output of plaintext in case --outfile-format 4, 5, 6 or 7 was used
- Fixed infinite loop when using --loopback in case all hashes have been cracked
- Fixed kernel loops in --increment mode leading to slower performance
- Fixed mask length check in hybrid attack-modes: Do not include hash-mode dependant mask length checks
- Fixed parsing of hashes in case the last line did not include a linefeed character
- Fixed potfile loading to accept blank passwords
- Fixed runtime limit: No longer required so sample startup time after refactorization

##
## Workarounds
##

- Workaround added for Intel OpenCL runtime: GPU support is broken, skip the device unless user forces to enable it

##
## Technical
##

- Building: Added hashcat32.dll and hashcat64.dll makefile targets for building hashcat windows libraries
- Building: Added production flag in Makefile to disable all the GCC compiler options needed only for development
- Building: Removed access to readlink() on FreeBSD
- Building: For CYGWIN prefer to use "opencl.dll" (installed by drivers) instead of optional "cygOpenCL-1.dll"
- Events: Added new event EVENT_WEAK_HASH_ALL_CRACKED if all hashes have been cracked during weak hash check
- Hardware management: Switched matching ADL device with OpenCL device by using PCI bus, device and function
- Hardware management: Switched matching NvAPI device with OpenCL device by using PCI bus, device and function
- Hardware management: Switched matching NVML device with OpenCL device by using PCI bus, device and function
- Hardware management: Switched matching xnvctrl device with OpenCL device by using PCI bus, device and function
- Hardware management: Removed *throttled* message from NVML as this created more confusion than it helped
- Hash Parser: Improved error detection of invalid hex characters where hex character are expected
- OpenCL Runtime: Updated AMDGPU-Pro driver version check, do warn if version 16.50 is detected which is known to be broken
- OpenCL Runtime: Updated hashcat.hctune for Iris Pro GPU on macOS
- Potfile: In v3.10 already, the default potfile suffix changed but the note about was missing. The "hashcat.pot" became "hashcat.potfile"
- Potfile: Added old potfile detection, show warning message
- Returncode: Added dedicated returncode (see docs/status_codes.txt) for shutdowns caused by --runtime and checkpoint keypress
- Sanity: Added sanity check to disallow --speed-only in combination with -i
- Sanity: Added sanity check to disallow --loopback in combination with --runtime
- Threads: Replaced all calls to ctime() with ctime_r() to ensure thread safety
- Threads: Replaced all calls to strerror() with %m printf() GNU extension to ensure thread safety

* changes v3.10 -> v3.20:

The hashcat core was completely refactored to be a MT-safe library (libhashcat).
The goal was to help developers include hashcat into distributed clients or GUI frontends.
The CLI (hashcat.bin or hashcat.exe) works as before but from a technical perspective it's a library frontend.

##
## Features
##

- New option --speed-only: Quickly provides cracking speed per device based on the user hashes and selected options, then quit
- New option --keep-guessing: Continue cracking hashes even after they have been cracked (to find collisions)
- New option --restore-file-path: Manually override the path to the restore file (useful if we want all session files in the same folder)
- New option --opencl-info: Show details about OpenCL compatible devices like an embedded clinfo tool (useful for bug reports)
- Documents: Added colors for warnings (yellow) and errors (red) instead of WARNING: and ERROR: prefix
- Documents: Added hints presented to the user about optimizing performance while hashcat is running
- Hardware management: Support --gpu-temp-retain for AMDGPU-Pro driver
- Hardware management: Support --powertune-enable for AMDGPU-Pro driver
- Password candidates: Allow words of length > 31 in wordlists for -a 0 for some slow hashes if no rules are in use
- Password candidates: Do not use $HEX[] if the password candidate is a valid UTF-8 string and print out as-is
- Pause mode: Allow quit program also if in pause mode
- Pause mode: Ignore runtime limit in pause mode
- Status view: Show core-clock, memory-clock and execution time in benchmark-mode in case --machine-readable is activated
- Status view: Show temperature, coreclock, memoryclock, fanspeed and pci-lanes for devices using AMDGPU-Pro driver
- Status view: Show the current first and last password candidate test queued for execution per device (as in JtR)
- Status view: Show the current position in the queue for both base and modifier (Example: Wordlist 2/5)
- Markov statistics: Update hashcat.hcstat which is used as reference whenever the user defines a mask
- Charsets: Added lowercase ascii hex (?h) and uppercase ascii hex (?H) as predefined charsets

##
## Algorithms
##

- Added hash-mode 14000 = DES (PT = $salt, key = $pass)
- Added hash-mode 14100 = 3DES (PT = $salt, key = $pass)
- Added hash-mode 14400 = SHA1(CX)
- Added hash-mode 99999 = Plaintext
- Extended hash-mode 3200 = bcrypt: Accept signature $2b$ (February 2014)
- Improved hash-mode 8300 = DNSSEC: Additional parsing error detection

##
## Bugs
##

- Custom charset from file parsing code did not return an error if an error occurred
- Fix some clSetKernelArg() size error that caused slow modes to not work anymore in -a 1 mode
- Hash-mode 11600 = (7-Zip): Depending on input hash a clEnqueueReadBuffer(): CL_INVALID_VALUE error occurred
- Hash-mode 22 = Juniper Netscreen/SSG (ScreenOS): Fix salt length for -m 22 in benchmark mode
- Hash-Mode 5500 = NetNTLMv1 + ESS: Fix loading of NetNTLMv1 + SSP hash
- Hash-mode 6000 = RipeMD160: Fix typo in array index number
- If cracking a hash-mode using unicode passwords, length check of a mask was not taking into account
- If cracking a large salted hashlist the wordlist reject code was too slow to handle it, leading to 0H/s
- Null-pointer dereference in outfile-check shutdown code when using --outfile-check-dir, leading to segfault
- On startup hashcat tried to access the folder defined in INSTALL_FOLDER, leading to segfault if that folder was not existing
- Random rules generator code used invalid parameter for memory copy function (M), leading to use of invalid rule
- Sanity check for --outfile-format was broken if used in combination with --show or --left

##
## Workarounds
##

- Workaround added for AMDGPU-Pro OpenCL runtime: Failed to compile hash-mode 10700 = PDF 1.7 Level 8
- Workaround added for AMDGPU-Pro OpenCL runtime: Failed to compile hash-mode 1800 = sha512crypt
- Workaround added for NVidia OpenCL runtime: Failed to compile hash-mode 6400 = AIX {ssha256}
- Workaround added for NVidia OpenCL runtime: Failed to compile hash-mode 6800 = Lastpass + Lastpass sniffed
- Workaround added for macOS OpenCL runtime: Failed to compile hash-mode 10420 = PDF 1.1 - 1.3 (Acrobat 2 - 4)
- Workaround added for macOS OpenCL runtime: Failed to compile hash-mode 1100 = Domain Cached Credentials (DCC), MS Cache
- Workaround added for macOS OpenCL runtime: Failed to compile hash-mode 13800 = Windows 8+ phone PIN/Password
- Workaround added for pocl OpenCL runtime: Failed to compile hash-mode 5800 = Android PIN

##
## Performance
##

- Improved performance for rule-based attacks for _very_ fast hashes like MD5 and NTLM by 30% or higher
- Improved performance for DEScrypt on AMD, from 373MH/s to 525MH/s
- Improved performance for raw DES-based algorithms (like LM) on AMD, from 1.6GH/s to 12.5GH/s
- Improved performance for raw SHA256-based algorithms using meet-in-the-middle optimization, reduces 7/64 steps
- Improved performance for SAP CODVN B (BCODE) and SAP CODVN F/G (PASSCODE) due to register handling optimization, gives 3% and 25%
- Improved performance by reducing maximum number of allowed function calls per rule from 255 to 31
- Improved performance by update the selection when to use #pragma unroll depending on OpenCL runtime vendor

- Full performance comparison sheet v3.10 vs. v3.20: https://docs.google.com/spreadsheets/d/1B1S_t1Z0KsqByH3pNkYUM-RCFMu860nlfSsYEqOoqco/edit#gid=1591672380

##
## Technical
##

- Autotune: Do not run any caching rounds in autotune in DEBUG mode if -n and -u are specified
- Bash completion: Removed some v2.01 leftovers in the bash completion configuration
- Benchmark: Do not control fan speed in benchmark mode
- Benchmark: On macOS, some hash-modes can't compile because of macOS OpenCL runtime. Skip them and move on to the next
- Building: Added Makefile target "main_shared", a small how-to-use libhashcat example
- Building: Added many additional compiler warning flags in Makefile to improve static code error detection
- Building: Added missing includes for FreeBSD
- Building: Added some types for windows only in case _BASETSD_H was not set
- Building: Changed Makefile to strip symbols in the linker instead of the compiler
- Building: Defined NOMINMAX macro to prevent definition min and max macros in stdlib header files
- Building: Enabled ASLR and DEP for Windows builds
- Building: Fixed almost all errors reported by cppcheck and scan-build
- Building: On macOS, move '-framework OpenCL' from CFLAGS to LDFLAGS
- Building: On macOS, use clang as default compiler
- Building: Support building on Msys2 environment
- Building: Use .gitmodules to simplify the OpenCL header dependency handling process
- Charsets: Added DES_full.charset
- Data Types: Replaced all integer macros with enumerator types
- Data Types: Replaced all integer variables with true bool variables in case they are used as a bool
- Data Types: Replaced all string macros with static const char types
- Data Types: Replaced all uint and uint32_t to u32
- Data Types: Replaced atoi() with atoll(). Eliminates sign conversion warnings
- Documents: Added docs/credits.txt
- Documents: Added docs/team.txt
- Documents: Changed rules.txt to match v3.20 limitations
- Error handling (file handling): Fixed a couple of filepointer leaks
- Error handling (format strings): Fixed a few printf() formats, ex: use %u instead of %d for uint32_t
- Error handling (memory allocation): Removed memory allocation checks, just print to stderr instead
- Error handling (startup): Added some missing returncode checks to get_exec_path()
- Fanspeed: Check both fanpolicy and fanspeed returncode and disable retain support if any of them fail
- Fanspeed: Minimum fanspeed for retain support increased to 33%, same as NV uses as default on windows
- Fanspeed: Reset PID controller settings to what they were initially
- Fanspeed: Set fan speed to default on quit
- File handling: Do a single write test (for files to be written later) directly on startup
- File locking: Use same locking mechanism in potfile as in outfile
- Hardware management: Fixed calling conventions for ADL, NvAPI and NVML on windows
- Hardware management: Improved checking for successful load of the NVML API
- Hardware management: In case fanspeed can not be set, disable --gpu-temp-retain automatically
- Hardware management: In case of initialization error show it only once to the user on startup
- Hardware management: Refactored all code to return returncode (0 or -1) instead of data for more easy error handling
- Hardware management: Refactored macros to real functions
- Hardware management: Removed kernel exec timeout detection on NVIDIA, should no longer occur due to autotune
- Hardware management: Replaced NVML registry functions macros with their ascii versions (Adds NVML support for XP)
- Hashlist loading: Do not load data from hashfile if hashfile changed during runtime
- Kernel cache: Fixed checksum building on oversized device version or driver version strings
- Logging: Improved variable names in hashcat.log
- Loopback: Refactored --loopback support completely, no longer a recursive function
- Memory management: Fixed some memory leaks on shutdown
- Memory management: Got rid of all global variables
- Memory management: Got rid of local_free() and global_free(), no longer required
- Memory management: Refactored all variables with HCBUFSIZ_LARGE size from stack to heap, macOS doesn't like that
- OpenCL Headers: Select OpenCL headers tagged for OpenCL 1.2, since we use -cl-std=CL1.2
- OpenCL Kernels: Added const qualifier to variable declaration of matching global memory objects
- OpenCL Kernels: Got rid of one global kernel_threads variable
- OpenCL Kernels: Moved OpenCL requirement from v1.1 to v1.2
- OpenCL Kernels: Recognize reqd_work_group_size() values from OpenCL kernels and use them in the host if possible
- OpenCL Kernels: Refactored common function append_0x01()
- OpenCL Kernels: Refactored common function append_0x02()
- OpenCL Kernels: Refactored common function append_0x80()
- OpenCL Kernels: Refactored rule function append_block1()
- OpenCL Kernels: Refactored rule function rule_op_mangle_delete_last()
- OpenCL Kernels: Refactored rule function rule_op_mangle_dupechar_last()
- OpenCL Kernels: Refactored rule function rule_op_mangle_rotate_left()
- OpenCL Kernels: Refactored rule function rule_op_mangle_rotate_right()
- OpenCL Kernels: Support mixed kernel thread count for mixed kernels in the same source file
- OpenCL Kernels: Switch from clz() to ffz() for bitsliced algorithms
- OpenCL Kernels: Using platform vendor name is better than using device vendor name for function detection
- OpenCL Runtime: Updated AMDGPU-Pro and AMD Radeon driver version check
- OpenCL Runtime: Updated Intel OpenCL runtime version check
- OpenCL Runtime: Updated NVIDIA driver version check
- Password candidates: The maximum word length in a wordlist is 31 not 32, because 0x80 will eventually be appended
- Potfile: Base logic switched; Assuming the potfile is larger than the hashlist it's better to load hashlist instead of potfile entries
- Potfile: In case all hashes were cracking using potfile abort and inform user
- Restore: Automatically unlink restore file if all hashes have been cracked
- Restore: Do not unlink restore file if restore is disabled
- Rules: Refactored macros to real functions
- Status: Added Input.Queue.Base and Input.Queue.Mod to help the user better understand this concept
- Status: Do not wait for the progress mutex to read and store speed timer
- Status: Do not show Recovered/Time when cracking < 1000 hashes
- Status: Do not show Recovered/Time as floats but as integers to reduce over-information
- Tests: Removed rules_test/ subproject: Would require total rewrite but not used in a long time
- Threads: Replaced all calls to getpwuid() with getpwuid_r() to ensure thread safety
- Threads: Replaced all calls to gmtime() with gmtime_r() to ensure thread safety
- Threads: Replaced all calls to strtok() with strtok_r() to ensure thread safety
- Wordlists: Use larger counter variable to handle larger wordlists (that is > 2^32 words)
- X11: Detect missing coolbits and added some help text for the user how to fix it

* changes v3.00 -> v3.10:

##
## Improvements
##

- Added mask display to modes 3, 6, and 7. Allows the user to see the custom character set used during the run
- Make Linux build POSIX compatible; Also allow it to actually compile on musl-libc systems
- Add support to compile on FreeBSD
- Make use of cl_context_properties[] to clCreateContext(), even if OpenCL specification allow the use of NULL, some runtimes fail without
- The Time.Estimated attribute in status display should also show --runtime limit if user set it
- Fix some strict aliasing rule violation on older compilers
- Fix some variable initializers on older compilers
- Replace DARWIN macro with compiler predefined macro __APPLE__
- Replace LINUX macro with compiler predefined macro __linux__
- Allow the use of enc_id == 0 in hash-mode 10600 and 10700 as it takes no part in the actual computation
- Get rid of exit() calls in OpenCL wrapper library with the goal to have a better control which error can be ignored under special circumstances
- Do not error and exit if an OpenCL platform has no devices, just print a warning and continue with the next platform
- Workaround for OpenCL runtimes which do not accept -I parameter in the OpenCL kernel build options even if this is an OpenCL standard option
- Workaround for OpenCL runtimes which do accept -I parameter in the OpenCL kernel build options, but do not allow quotes
- Output cracked hashes on Windows using \r\n and not \n
- Replace RegGetValue() with RegQueryValueEx() to enable Windows XP 32 bit compatibility
- Slightly increased NVidias rule-processing performance by using generic instructions instead of byte_perm()
- Add support for @ rule (RULE_OP_MANGLE_PURGECHAR) to use on GPU
- Add support for --outfile (short -o) to be used together with --stdout
- Skip periodic status output whenever --stdout is used together with stdin mode, but no outfile was specified
- Show error message if --show is used together with --outfile-autohex-disable (this is currently not supported)
- Show error message if --skip/--limit is used together with mask files or --increment
- Workaround for NVidia OpenCL runtime bug causing -m 6223 to not crack any hashes even with the correct password candidate

##
## Bugs
##

- Fixed a bug where CRAM MD5 checked salt length instead of hash length
- Fixed a bug where hashcat is suppressing --machine-readable output in the final status update
- Fixed a bug where hashcat did not check the return of realpath() and crashes uncontrolled if the path does not exist
- Fixed a bug where hashcat crashes for accessing deallocated buffer if user spams "s" shortly before hashcat shuts down
- Fixed a bug where hashcat crashes in case of a scrypt P setting > 1
- Fixed a bug where hashcat did not correctly use the newly cracked plains whenever --loopback or the induction folder was used
- Fixed a bug where hashcat did not correctly remove hashes of type WPA/WPA2 even if present in potfile
- Fixed a bug where hashcat reported an invalid password for a zero-length password in LM
- Fixed a bug where hashcat did not take into account how long it takes to prepare a session when auto-aborting with --runtime is in use
- Fixed a bug where some kernels used COMPARE_M_SIMD instead of COMPARE_S_SIMD in singlehash mode

##
## Algorithms
##

- Added new hash-mode 13900 = OpenCart

* changes v2.01 -> v3.00:

This release marks the fusion of "hashcat" and "oclHashcat" into "hashcat".
It combines all features of all hashcat projects in one project.

##
## Features
##

- Support for Apple OpenCL runtime
- Support for NVidia OpenCL runtime (replaces CUDA)
- Support for Mesa (Gallium) OpenCL runtime
- Support for pocl OpenCL runtime
- Support for Khronos' OSS OpenCL reference implementation for building
- Support to utilize OpenCL devices-types other than GPU, ex: CPU and FPGA
- Support to utilize multiple different OpenCL platforms in parallel, ex: AMD + NV
- Support to utilize multiple different OpenCL device-types in parallel, ex: GPU + CPU
- Added option --opencl-platform to select a specific OpenCL platform
- Added option --opencl-device-types select specific OpenCL device types
- Added option --opencl-vector-width to override automatically selected vector-width size
- Added makefile native compilation target
- Added makefile install and uninstall targets
- Added autotuning engine and user-configurable tuning database
- Added current engine clock, current memory clock and pci-e lanes to the status display
- Added support for --gpu-temp-retain for NVidia GPU, both Linux and Windows
- Added execution timer of the running kernel to the status display
- Added command prompt to quit at next restore checkpoint
- Added human-readable error message for the OpenCL error codes
- Added option --potfile-path to override potfile path
- Added option --veracrypt-keyfile to set Keyfiles used, can be multiple
- Added option --veracrypt-pim to set the VeraCrypt personal iterations multiplier
- Added option --machine-readable for easier parsing of output
- Added option --powertune-enable to work with NVidia devices as well, not just AMD
- Added option --stdout to print candidates instead of trying to crack a hash

##
## Algorithms
##

- Added new hash-mode   125 = ArubaOS
- Added new hash-mode 12900 = Android FDE (Samsung DEK)
- Added new hash-mode 13000 = RAR5
- Added new hash-mode 13100 = Kerberos 5 TGS-REP etype 23
- Added new hash-mode 13200 = AxCrypt
- Added new hash-mode 13300 = AxCrypt in memory SHA1
- Added new hash-mode 13400 = Keepass 1 (AES/Twofish) and Keepass 2 (AES)
- Added new hash-mode 13500 = PeopleSoft PS_TOKEN
- Added new hash-mode 13600 = WinZip
- Added new hash-mode 137** = VeraCrypt
- Added new hash-mode 13800 = Windows 8+ phone PIN/Password

##
## Performance
##

- Full Table: https://docs.google.com/spreadsheets/d/1B1S_t1Z0KsqByH3pNkYUM-RCFMu860nlfSsYEqOoqco/edit#gid=0

##
## Improvements
##

- Reordering of files to help integration into linux distributions ~/.hashcat etc
- Use a profile directory to write temporary files (session, potfile etc.)
- Workaround dependencies on AMD APP-SDK AMD ADL, NV CUDA-SDK, NV ForceWare, NVML and NVAPI; they are no longer required
- Load external libraries dynamic at runtime instead of link them static at compile-time
- Benchmark accuracy improved; Is now on par to: singlehash -a 3 -w 3 ?b?b?b?b?b?b?b
- Benchmark no longer depends on a fixed time
- Removed option --benchmark-mode, therefore support --workload-profile in benchmark-mode
- Enabled support of --machine-readable in combination with --benchmark for automated benchmark processing
- Replaced --status-automat entirely with --machine-readable to make it more consistent among benchmark and non-benchmark mode
- Extended support from 14 to 255 functions calls per rule
- Extended password length up to 32 for 7zip
- Extended salt length up to 55 for raw hash types, eg: md5($pass.$salt)
- Extended version information
- Removed some duplicate rules in T0XlCv1, d3ad0ne and dive
- Redesigned changes.txt layout
- Redesigned --help menu layout

##
## Bugs
##

- Fixed a bug in speed display: In some situation, especially with slow hashes or lots of salts, it showed a speed of 0H/s
- Fixed a bug in restore handling: user immediately aborting after restart broke the restore file
- Fixed a bug in line counter: conditional jump or move depends on an uninitialised value
- Fixed a bug in rule-engine for NVidia devices: code for left- and right-shift were switched
- Fixed a bug in dive.rule: rules were not updated after the function 'x' was renamed to 'O'
- Fixed a bug in memory allocation "OpenCL -4 error": used uninitialized value in a special situation
- Fixed a bug in memory handling: heap buffer overflow
- Fixed a bug in memory handling: out of bounds access
- Fixed a bug in implementation of DCC2: forced default iteration count for hashes to 10240
- Fixed a bug in implementation of WPA/WPA2: MAC and nonce stay one their original position as in the hccap file
- Fixed a bug in implementation of GOST R 34.11-94: zero length passwords were not cracked
- Fixed a bug in implementation of BLAKE2-512 kernels: incorrect access of the esalt buffer

##
## Technical
##

- Removed deprecated GCC version check requirement
- Removed NPROCS from Makefile, let make automatically detect the optimal number of parallel threads
- Dropped all C++ overloading functions to normal function which helps support more OpenCL platforms
- Renamed functions in common.h to emphasize their purpose
- Refactorized fast-hash kernels to enable SIMD on all OpenCL platforms
- Refactorized SIMD handling: SIMD the inner-loop not the outer-loop to save registers
- Workaround missing clEnqueueFillBuffer() support in certain OpenCL runtimes
- Added amd_bytealign() support in non-AMD OpenCL runtimes
- Added amd_bfe() support in non-AMD OpenCL runtimes
- Added several macros to allow writing optimized code for the different OpenCL platforms
- Replaced typedef for bool with stdbool.h
- Added special DEBUG environment variables to the makefile
- Hashcat now acquires an exclusive lock before writing to any file
- Changed buffers to not use same buffer for both input and output at the same time with snprintf()
- Check for allocatable device-memory depending on kernel_accel amplifier before trying to allocate
- Added additional check for max. ESSID length to prevent possible crashes
- Use a GCC equivalent for __stdcall where applicable
- Synchronize maximum output line size with input line size
- Increased maximum hash line size to 0x50000
- Run weak-hash checks only in straight-attack mode, this greatly reduces code complexity
- Restrict loopback option to straight attack-mode
- Moved rules_optimize to hashcat-utils
- Stick to older libOpenCL in binary package to avoid errors like this: version `OPENCL_2.0' not found
- Tightened hash parser for several algorithms
- Updated old RC4 code in Kerberos 5
- Limited the salt length of Juniper Netscreen/SSG (ScreenOS) hashes to 10
- Updated algorithm used to automatically select an ideal --scrypt-tmto value
- Renamed option --gpu-accel to --kernel-accel
- Renamed option --gpu-loops to --kernel-loops
- Renamed option --gpu-devices to --opencl-devices
- Added inline declaration to functions from simd.c, common.c, rp.c and types_ocl.c to increase performance
- Dropped static declaration from functions in all kernel to achieve OpenCL 1.1 compatibility
- Added -cl-std=CL1.1 to all kernel build options
- Created environment variable to inform NVidia OpenCL runtime to not create its own kernel cache
- Created environment variable to inform pocl OpenCL runtime to not create its own kernel cache
- Dropped special 64-bit rotate() handling for NV, it seems that they've added it to their OpenCL runtime
- Completely get rid of HAVE_ADL, HAVE_NVML and HAVE_NVAPI in sources
- Replaced NVAPI with NVML on windows
