hate_crack/CHANGELOG.md

Changelog

All notable changes to this project are documented here.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Dates are omitted for releases predating this file; see the git tags for exact timing.

[2.14.3] - 2026-07-25

Added

• Private-key commit gate. prek.toml now runs the detect-private-key hook at the pre-commit stage. The repo previously had no secret-scanning gate of any kind — bandit only covers hate_crack/, so nothing inspected config files, docs, or test fixtures for committed key material.

Removed

• Local agent tooling is no longer published. CLAUDE.md, .claude/, docs/plans/, and docs/superpowers/ were development aids rather than part of the shipped project. They are now gitignored and were removed from the repository, including from its history.
• audit-docs post-commit hook. Dropped from prek.toml along with the .claude/audit-docs.sh script it invoked.

[2.14.2] - 2026-07-25

Fixed

• Pipal base-word parsing. pipal() built one rigid regex that required exactly pipal_count consecutive base-word lines, so any cracked set with fewer unique base words than pipal_count (default 10 — the common case on small cracks) matched nothing and returned no base words. The Top N base words section is now parsed line by line, returning up to pipal_count words and stopping at the end of the section.
• Shell-safe pipal invocation. The pipal subprocess is now spawned with list-form arguments instead of a shell=True formatted string, so hash-file paths containing shell metacharacters can no longer be interpreted as commands.

Changed

• Renamed the internal _omen_pick_training_wordlist helper to _pick_training_wordlist, since it is shared by the OMEN, Markov-adjacent, and LLM (wordlist mode) attacks rather than being OMEN-specific.

[2.14.1] - 2026-07-25

Fixed

• Tab completion on custom file-path prompts. The p. Enter a custom path branches of the OMEN and Markov training pickers, the combipow wordlist prompt, and the rule cleanup/optimize output-path prompts used a bare input() with no readline completer, so TAB did nothing. They now route through select_file_with_autocomplete for consistent path autocompletion.
• Stale completer leak. select_file_with_autocomplete and the _configure_readline-based pickers now drop the path completer after a selection, so later numeric-menu and y/n prompts no longer inherit file-path tab completion.

[2.14.0] - 2026-07-24

Added

• Non-interactive attack subcommands for scripting (issue #17). Launch a single attack without the menu: quick (wordlist + optional --rules), dict (configured-wordlist methodology), brute (--min/--max), and topmask (--target-time). Preprocessing prompts auto-accept their defaults, and the process returns a clean exit code (0 on success, non-zero on a bad hash file, hash type, wordlist, or rule name).

[2.13.1] - 2026-07-24

Fixed

• OLLAMA_HOST values that include a scheme no longer produce a malformed URL (issue #119). The Ollama base URL was built as "http://" + OLLAMA_HOST, so a value in the form Ollama's own tooling accepts — http://box:11434 or https://ollama.example.com — became http://http://box:11434 and the LLM attack could not connect. http:// is now prepended only when no scheme is present, and trailing slashes are stripped because callers append paths (f"{ollamaUrl}/v1"). Reaching a remote Ollama over TLS works as a result. The bare host:port default is unchanged.

[2.13.0] - 2026-07-24

Added

• Cracked-password generation mode for the LLM attack. Once a session has recovered plaintexts, option 3 feeds them back to the model, which infers the organization's own password conventions and generates new candidates in that style. Offered only when <hashfile>.out has content, and it uses a dedicated prompt that tells the model not to re-emit passwords already cracked.
• Target research pre-fills the industry and location prompts. In target mode, entering the company name asks the local model to recall that organization's industry and location, then offers them as editable defaults (Enter accepts, typing overrides). Values are labelled as model guesses rather than verified OSINT, whitespace-collapsed, and capped at 80 characters. Research runs entirely against the local Ollama server, so the client name is never sent to a third party. Any failure or timeout falls back to blank prompts and never blocks the attack. Disable with ollamaAutoResearch: false.
• Live progress spinner with an elapsed-seconds counter during Ollama generation, so a model loading into VRAM is distinguishable from a hang. Automatically suppressed when stdout is not a TTY.
• ollamaMaxSampleLines (default 500) caps how many sample passwords are sent to the model, for both wordlist and cracked-password modes.

Fixed

• A large sample wordlist no longer stalls the LLM attack. Wordlist mode read every line into memory and pasted all of them into the prompt, so pointing it at rockyou.txt materialized hundreds of megabytes and overran the model's context window — which looked like a hang. The file is now streamed and evenly sampled across its whole length, and the count actually used is reported (Sampled 500 of 14,344,391 passwords from wordlist.).
• HATE_CRACK_ARROW_MENU=1 now works in the LLM and OMEN submenus. They hand-rolled print() + input() instead of the shared menu helper, so arrow-key navigation silently did nothing there.
• A typo in a wordlist or generation-mode prompt no longer aborts the whole attack. The pickers and submenus re-prompt instead of dropping back to the main menu, and offer an explicit cancel.

Changed

• Interactive prompt formatting normalized — the [*] marker is no longer used on input prompts (it denotes status output elsewhere), and default-value hints use a single form.

Build

• Local uv Python pinned to 3.13 via .python-version. requires-python = ">=3.13" meant a fresh worktree picked CPython 3.15.0a7 and failed to build pyo3 0.26 (via jiter/fastuuid/pydantic-core). CI already pinned 3.13.

[2.12.0] - 2026-07-24

Changed

• LLM attack now uses the Atomic Agents framework for structured (JSON) candidate generation instead of raw HTTP + regex line-parsing. Candidate generation lives in the new hate_crack/llm.py module.
• Default Ollama model is now qwen2.5:32b (was mistral), chosen for reliable structured-output adherence.

Added

• Wordlist (denylist) generation mode for the LLM attack is now reachable from the menu: select the LLM attack (option 12), then choose "Wordlist" to derive basewords from a sample wordlist.

Fixed

• The LLM attack no longer hangs forever waiting on Ollama. Generation requests are now bounded by a configurable timeout (ollamaTimeout in config.json, default 300 seconds). Previously, if Ollama accepted the connection but never replied — most commonly a large model still loading into VRAM — hate_crack sat at a frozen prompt with no recourse but Ctrl-C. When the timeout fires you now get a specific message naming the elapsed timeout and the setting to raise, instead of a misleading "ensure Ollama is running" hint.

Removed

• Automatic model pulling. hate_crack no longer pulls missing Ollama models; pull them yourself with ollama pull <model>.

[2.11.4] - 2026-07-24

Added

• CI now runs lint, type checks, and tests on every pull request and push to main. Previously these ran only in local prek pre-push hooks, so a commit pushed without hooks installed (or with --no-verify) reached main unvalidated — and the auto-tag workflow would then cut a release from it. The new ci.yml runs ruff, ty, and pytest on Python 3.13, and tagging is gated on it passing.
• Dependabot configuration for Python (uv) and GitHub Actions dependencies, weekly. Action pins are updated automatically instead of drifting. Bumps use chore commit prefixes so they never auto-cut a release.

Fixed

• Auto-tagged versions now actually get a GitHub release. The tag was pushed using the default GITHUB_TOKEN, and GitHub suppresses workflow triggers for GITHUB_TOKEN-created events, so the tag-triggered release workflow never fired — v2.11.3 was tagged with no release to show for it. The auto-tag workflow now creates the release itself, idempotently.
• Breaking changes no longer ship as a minor bump. The version logic matched the ! breaking-change marker but only ever bumped the minor version, and BREAKING CHANGE: footers were invisible because only commit subjects were inspected. Both now correctly trigger a major bump.
• Concurrent merges to main no longer collide. Two merges in quick succession both computed the same new tag and the second push failed; tagging is now serialized. A re-run against an already-tagged commit is also a clean no-op instead of a hard error.
• softprops/action-gh-release is pinned to a real release (v2.6.2) rather than an arbitrary mid-development master commit, and all workflows now pin the same actions/checkout version with accurate version comments.
• ty type error in the crack tailer. _read_new_lines read self._file_pos (int | None) while its only None-guard lived in the caller, so the position is now passed in explicitly as an int. Behavior is unchanged.

[2.11.3] - 2026-07-24

Fixed

• Rule file cleanup no longer errors out. rules_cleanup (Rule File Tools → "Clean rule file" / "Clean and optimize") invoked cleanup-rules.bin with no arguments, but the binary requires a mode argument (1 = CPU, 2 = GPU) and exits with usage text otherwise — so cleanup always failed. It now passes the mode (defaulting to GPU) so the cleanup actually runs.

[2.11.2] - 2026-07-23

Added

• Hashview cracked-hash upload now reports how many hashes landed. After an upload the CLI prints the number of pairs the client sent and how many it skipped by validation — available regardless of the server version — and, against a Hashview that reports import counts, also shows how many were newly cracked, verified, and left unmatched (already cracked or not present). upload_cracked_hashes surfaces uploaded/skipped in its return value alongside any server-provided counts. Previously the upload only printed ✓ Success: OK with no indication of how many hashes were accepted.

[2.11.1] - 2026-07-23

Fixed

• Hashview cracked-hash uploads no longer choke on hashcat $HEX[...] plaintexts. hashcat emits $HEX[...] for recovered passwords containing leading/trailing whitespace or non-UTF-8 bytes. upload_cracked_hashes forwarded those verbatim, so a Hashview that verifies the plaintext against the hash rejected the entire batch with Plaintext for hash ... was found to be invalid. The uploader now decodes $HEX[...] to the exact bytes the server must re-hash — latin-1→UTF-8 for the UTF-16LE modes (NTLM 1000, MSSQL 1731), raw bytes for the raw-byte modes (0/100/300/900/1400/1700) — and keeps the $HEX wrapper verbatim when inlining would be unsafe (embedded CR/LF) so a $HEX-aware server can still handle it. Verified end-to-end against an unpatched Hashview.

Added

• Client-side validation of cracked hash:plaintext pairs before Hashview upload. upload_cracked_hashes now filters each pair against the declared hashcat mode: a length check for wrong-width hashes, plus a plaintext recompute for the reproducible fast modes (MD5, SHA1, MD4, NTLM, SHA2-256/512). Mismatched lines (e.g. a stray MD5 hash mixed into an NTLM list) are skipped with a per-line warning instead of failing the whole upload server-side, and it raises clearly if nothing valid remains. Bundles a pure-Python MD4 since OpenSSL 3 dropped it from hashlib. Opt out with validate=False.

[2.11.0] - 2026-07-23

Changed

• Shard Wordlist (Wordlist Tools option 7) now produces all shards in a single run. Instead of prompting for a modulus + offset and emitting one file per invocation, it prompts for an output base path and a shard count (N), then writes all N interleaved parts named with zero-padded part numbers (base.001…base.00N). This matches the intended distributed-cracking workflow — split once, copy one part per node — without re-running the tool for each offset. README usage docs updated accordingly.

[2.10.10] - 2026-07-21

Security

• Bumped the pinned dev/test dependency pytest from 9.0.2 to 9.0.3 to clear the vulnerable tmpdir-handling advisory GHSA-6w46-j5rx-g56g (affects pytest < 9.0.3; Dependabot alert #1). Development-scope test runner only — no runtime dependency change. The full test suite passes under 9.0.3. (uv.lock is gitignored, so only the pyproject.toml pin is tracked.)

[2.10.9] - 2026-07-21

Fixed

• Quick Crack default wordlist stays on hcatOptimizedWordlists. The numbered list and tab-completion browse hcatWordlists, but pressing Enter still falls back to hcatOptimizedWordlists as before.

[2.10.8] - 2026-07-21

Fixed

• Hashview list_customers crashed against current servers. The /v1/customers response now returns its users array as native JSON (Hashview issue #229), but the client still ran json.loads() on it unconditionally, raising TypeError and breaking the entire customer → hashfile enumeration flow. Both the native-array and the legacy double-encoded-string shapes are now accepted.
• Hashview hash-type parsing mis-read MD5 (mode 0). get_hashfile_details selected the hash type with an or fallthrough, so the falsy 0 fell through to the response envelope's type field and returned the string "message". Hash type is now read by key presence, and the bogus type fallback was removed.
• Hashview get_hashfile_hash_type always returned an empty list. It looked for file_ids/ids/hashfile_ids keys the endpoint never sends; it now reads the actual hashfiles envelope array and extracts each file id.

Added

• Download Hashview rule files. New HashviewAPI.list_rules() and download_rules() wrap GET /v1/rules and GET /v1/rules/{id}. The server gzip-compresses plaintext rules on the fly, so downloads are decompressed before saving — the resulting file is usable directly with hashcat -r. Exposed via the interactive Hashview menu ("Download Rule") and the CLI: hate_crack.py --hashview download-rules --rules-id <id> [--output <file>].

[2.10.7]

• Auto-upgrade (hate_crack --update / the in-menu upgrade) now survives the historical master → main default-branch rename. Old clones made before the rename sit on a local master whose upstream (branch.master.merge) still points at the now-deleted refs/heads/master, so a bare git pull failed with "Your configuration specifies to merge with the ref 'refs/heads/master' from the remote, but no such ref was fetched" — and _run_upgrade()'s git checkout main also failed on stale clones that had never fetched origin/main. The updater now fetches from origin before switching branches, checks out main via git checkout -B main origin/main (creating/resetting it from the remote regardless of local state), repairs the upstream with git branch --set-upstream-to=origin/main main so future manual git pulls work too, and pulls explicitly with git pull origin main so it never consults the dangling branch.*.merge config. Existing safety guards (dirty-branch bail, detached-HEAD skip) are unchanged.

[2.10.6]

• Fixed the Hashview integration calling API routes that don't exist in Hashview (verified against v0.8.3-dev), which 404'd as soon as a customer ID was entered ("Could not list hashfiles"). The customer→hashfile listing relied on a phantom GET /v1/hashfiles list-all route; it now enumerates via the real GET /v1/hashfiles/hash_type/<type> endpoint where available — the download flow sweeps common hashcat modes to display a customer's uploaded hashfiles. That listing route only exists on Hashview builds from 2026-06-08+ (the v0.8.3-dev branch); on main/older servers there is no hashfile-listing API at all, so the flow now degrades gracefully to entering the hashfile ID directly (looked up in the Hashview web UI) and resolving its type via GET /v1/getHashType/<id>. Additional client-side route fixes: hashfile hash-type lookup now uses GET /v1/getHashType/<id>; "left" (uncracked) hash download uses GET /v1/hashfiles/<id>; delete_job uses DELETE /v1/jobs/<id>; start_job uses POST. Hashview exposes no stop-job route, so stop_job now raises with guidance to use delete_job; and no bulk cracked-hash export exists, so the best-effort "found" merge degrades gracefully.
• The Hashview CLI now honours HASHVIEW_URL / HASHVIEW_API_KEY environment variables as overrides for the config.json values, so the client can be pointed at a different Hashview instance (e.g. a local dev stack) without editing the persisted config.
• Added an opt-in local Hashview integration-test harness: HASHVIEW_TEST_LOCAL=1 (with HASHVIEW_REPO=<path>) spins up and seeds a local Hashview docker stack, runs the live Hashview tests against it, and tears it down (HASHVIEW_KEEP=1 keeps it). This is what surfaced and verified the route fixes above against v0.8.3-dev. See the README testing section for details.

[2.10.5]

• Pipal analysis no longer corrupts its input when cracked passwords contain $HEX[...] rows. binascii.unhexlify().decode() returned the bytes without the trailing newline that normal rows inherit from password[-1], so every HEX-encoded password got concatenated with the next one in the .passwords file fed to pipal (e.g. three cracks → two lines, one of them a bogus mashup). Pipal then under-counted entries and reported wrong top base words. The HEX branch now re-appends \n so each cracked password lands on its own line.

[2.10.4]

• Pushover notifications fire correctly for Quick Crack, Loopback, Combinator, PRINCE-LING, and N-gram attacks (#110). The handlers prompted the user under one name (e.g. "Quick Crack") while the underlying hashcat wrapper passed a different attack_name to _should_fire ("Quick Dictionary"), so the per-run consent lookup always missed. The prompt name now flows down to _run_hcat_cmd for both the job-done summary and the per-crack tailer.

[2.10.3]

• Auto-upgrade no longer loops infinitely when invoked from a non-main branch (e.g. dev). Release tags live on main-side merge commits, so git pull on dev was a no-op and setuptools-scm kept regenerating the version as X.Y.Z.postN.devM — the update check then re-fired forever. _run_upgrade() now switches to main before pulling, with safety guards: refuses to clobber uncommitted work, surfaces clear errors when main is checked out in another worktree, and leaves detached-HEAD checkouts untouched.

[2.10.2]

• Fingerprint Attack no longer launches hashcat against empty wordlists when no candidates exist; prints a "no candidates to expand" message and skips the attack (plus the secondary hybrid pass that previously fired six wasted hashcat sessions).
• Forced LC_ALL=C on every sort -u subprocess (fingerprint expander pipeline, _write_field_sorted_unique, LM-to-NT combinator dedupe) — fixes "sort: Illegal byte sequence" on macOS when cracked passwords contain non-UTF-8 bytes, which was silently emptying the fingerprint candidate list.

[2.10.1]

• Bumped HashcatRosetta submodule to v0.2.0, dropping a vulnerable transitive pytest (< 9.0.3, GHSA tmpdir handling) from its requirements.
• Added click>=8.0.0 to runtime dependencies (now required by HashcatRosetta v0.2.0's formatting module).

[2.9.3]

• Transmission daemon now watches /tmp/hate_crack/ for new .torrent files; wordlist content still downloads to the configured wordlist directory.
• Suppressed transmission-daemon stdout/stderr so daemon log output no longer appears in the terminal.
• Increased watch-dir polling window to 30s to account for transmission's ~10s scan interval.
• Store downloaded .torrent files in /tmp/hate_crack/ instead of /tmp/ root.

[2.5.0]

• Added tab autocomplete to all file and directory path prompts in the Wordlist Tools submenu (option 80).
• Restored hcatOptimizedWordlists config key (directory for pre-optimized wordlists); defaults to ./optimized_wordlists, falls back to hcatWordlists if not found.
• Quick Crack now defaults to hcatOptimizedWordlists instead of hcatWordlists.

[2.0+]

• Added Random Rules Attack (option 20) using generate-rules.bin to generate random mutation rules (#87).
• Added Ad-hoc Mask Attack (option 17) for user-typed hashcat masks with optional custom character sets.
• Added Markov Brute Force Attack (option 18) using hcstat2 statistical tables for password generation.
• Consolidated Combinator Attacks (formerly options 10/11/12) into interactive submenu under option 6.
• Markov attack supports training from cracked passwords or any wordlist, with table reuse/regeneration menu.
• Fixed OMEN attack failing silently when model files were incomplete or enumNG errors occurred.
• OMEN attack now validates all 5 required model files, captures enumNG stderr, and provides a train/use/cancel menu with wordlist picker.
• Filtered .7z, .torrent, and .out files from wordlist selection menus (#80).
• Parallelized Hashmob rule downloads using a thread pool with success/failure summary (#81).
• Added dynamic optimized kernel (-O) flag per attack type via optimizedKernelAttacks config (#82).
• Replaced uv tool install with a bash shim for reliable config and asset resolution from any working directory.
• Fixed config resolution to search the repo root and package directory in addition to CWD.
• Fixed bare NTLM hash detection failing when hash files contain leading blank lines, BOM characters, or null bytes from UTF-16 encoding.
• Improved error message for unrecognized hash formats to show the actual first-line content and list expected formats.
• Fixed rule file path construction in Quick Crack and Loopback Attack using os.path.join() instead of string concatenation.
• Added automatic update checks on startup (check_for_updates config option).
• Added packaging dependency for version comparison.
• Added OMEN Attack (option 16) using statistical model-based password generation.
• Added OMEN configuration keys (omenTrainingList, omenMaxCandidates).
• Added LLM Attack (option 15) using Ollama for AI-generated password candidates.
• Added Ollama configuration keys (ollamaModel, ollamaNumCtx).
• Auto-versioning via setuptools-scm from git tags.

[2.0]

• Modularized codebase into CLI/API/attacks modules.
• Unified CLI options with config overrides (hashview, hashcat, wordlists, pipal).
• Added Hashview API integration.
• Added Weakpass torrent download helpers and Hashmob download wrapper.
• Improved test coverage and snapshot-based menu validation.
• Updated documentation and versioning.

[1.9]

• Revamped the hate_crack output to increase processing speed exponentially; combine_ntlm_output function for combining.
• Introducing new attack mode "Bandrel Methodology".
• Updated pipal function to output top x number of basewords.

[1.08]

• Added a Pipal menu option to analyze hashes. https://github.com/digininja/pipal

[1.07]

• Minor bug fixes with pwdump formatting and unhexify function.

[1.06]

• Updated the quick crack and recycling functions to use user-customizable rules.

[1.05]

• Abstraction of rockyou.txt so that you can use whatever dictionary you would like, specified in the config.json.
• Minor change to quickcrack that allows you to specify 0 for the number of times best64 is chained.

[1.04]

• Two new attacks: Middle Combinator and Thorough Combinator.

[1.03]

• Introduction of new feature to use session files for multiple concurrent sessions of hate_crack.
• Minor bug fix.

[1.02]

• Introduction of new feature to export the output of pwdump-formatted NTDS outputs to Excel with clear-text passwords.

[1.01]

• Minor bug fixes.

[1.00]

• Initial public release.