pcfg-go - The Fast PCFG

[cyclone]

Title: pcfg-go
Author: cyclone
URL: https://github.com/cyclone-github/pcfg-go
Description: Probabilistic Context Free Grammar (PCFG) password generator in Pure Go

pcfg-go - The Fast PCFG

• Probabilistic Context-Free Grammar (PCFG) password generator - Pure Go Edition
  • pcfg-go is a Pure Go rewrite of the Python3 pcfg_cracker
  • The goal of this Go implementation is to provide a substantial performance improvement for both trainer and guesser over the original Python3 version, while also adding features such as supporting $HEX[] input and multi-byte character support — which is not implemented in the Pure C pcfg_guesser
  • Credits for the original python3 pcfg_cracker belong to the author @lakiw
  • Rules trained by pcfg_trainer are compatible with hashcat's new PCFG -a 4 attack mode. See the hashcat PCFG implementation commit for details.

---

Install

pcfg_trainer:

go install -ldflags="-s -w" github.com/cyclone-github/pcfg-go/cmd/pcfg_trainer@main

pcfg_guesser:

go install -ldflags="-s -w" github.com/cyclone-github/pcfg-go/cmd/pcfg_guesser@main

---

Additions & improvements

• Performance — ~53.94× faster pcfg_trainer and ~12.16× faster pcfg_guesser vs Python 3 (see Benchmarks below)
• $HEX[] input — Trainer accepts $HEX encoded passwords in the training wordlist (multi-byte support)
• Ctrl+C handling — Pressing Ctrl+C auto-saves the session in pcfg_guesser
• Multi-keyboard layouts — QWERTY, AZERTY, QWERTZ, Dvorak, JCUKEN (Russian Cyrillic)
• Expanded TLD list — Legacy TLDs, ccTLDs, gTLDs (.info, .xyz, .app, .dev, etc.), and short TLDs (.co, .io, .ai, .me, .gg); improves both website and email detection
• Improved website detection — Broader URL/prefix detection (http://, https://, www., etc.) and host extraction
• Multi-threaded architecture — pcfg_guesser is multi-threaded for increased performance
• Compiled binary — No fuss; pcfg-go uses compiled binaries for speed and easy deployment

---

Benchmarks on rockyou training set

pcfg_trainer -r rockyou -t rockyou.txt

• Python3 trainer: 13m21s (801s)
• Go pcfg_trainer: 14.85s
• Speedup: ~53.94×

pcfg_guesser -r rockyou -n 100000000 (100 million guesses written to stdout)

• Python3 pcfg_guesser: 195.5s
• Go pcfg_guesser: 16.08s
• Speedup: ~12.16×

---

Usage

pcfg_trainer

Train a new ruleset from wordlist:

pcfg_trainer -r rule_name -t wordlist.txt

pcfg_guesser

Generate guesses from a trained ruleset:

pcfg_guesser -r rule_name

Session save/restore:

pcfg_guesser -r rule_name -s my_session   # save to my_session.sav on exit
pcfg_guesser -r rule_name -s my_session -l # load and resume

Press Ctrl+C to save session and exit.

Piping into hashcat

pcfg_guesser -r rule_name -s my_session | hashcat -m 0 hashes.txt...

Using trained rules with hashcat -a 4 PCFG mode

hashcat -m 0 -a 4 hashes.txt path_to_pcfg_rule_dir ...

---

Flags

pcfg_trainer

pcfg-go vs pcfg-python3 flags

Go	Python3	Description
-r	--rule	Ruleset name
-t	--training	Training wordlist (required)
-e	--encoding	File encoding
-C	--comments	Config comments
-S	--save_sensitive	Save emails, URLs
-p	--prefixcount	Lines prefixed with count
-n	--ngram	OMEN ngram size (2-5)
-a	--alphabet	Alphabet size for Markov
-c	--coverage	PCFG vs OMEN coverage
-m	--multiword	Pre-train multiword file
-h	--help	Help
-version	--version	Version info

pcfg_guesser

pcfg-go vs pcfg-python3 flags

Go	Python3	Description
-r	--rule	Ruleset name
-s	--session	Session name
-l	--load	Load previous session
-n	--limit	Max guesses
-b	--skip_brute	Skip OMEN/Markov
-a	--all_lower	No case mangling
-d	--debug	Debug output
-h	--help	Help
-version	--version	Version info

[cyclone]

https://github.com/cyclone-github/pcfg-go/releases/tag/v0.5.0

v0.5.0; 2026-03-16

initial github release

### Overview
Pure Go rewrite of the Python3 `pcfg_cracker`, designed as a near drop-in replacement with significant performance gains and expanded features

### Highlights
~3× faster trainer
~40× faster pcfg_guesser
`$HEX[]` input/output support
Full multi-byte / Unicode support (not supported in Compiled C Edition)
Improved and expanded keyboard detection: Fixed/tuned: QWERTY, JCUKEN Added: AZERTY, QWERTZ, Dvorak
Expanded detection for TLDs, URLs, and emails in trainer
Auto-save and resume support in pcfg_guesser
Multi-threaded for high-throughput performance

[cyclone]

v0.5.1; 2026-03-17
https://github.com/cyclone-github/pcfg-go/releases/tag/v0.5.1

fix: unicode handling
tweak: session save/restore

[cyclone]

v0.5.2; 2026-03-23
https://github.com/cyclone-github/pcfg-go/releases/tag/v0.5.2

Fixed Issue #5 -n {nth} not outputting

[cyclone]

v0.5.3; 2026-07-25
https://github.com/cyclone-github/pcfg-go/releases/tag/v0.5.3

• pcfg_guesser ~2× faster than v0.5.2
• Reduce long-run RAM growth in pcfg_guesser (priority-queue frontier)
  • Compact parse-tree representation (interned type IDs, packed nodes, arena reuse)
  • Contiguous index-based heap (fewer per-item allocations)
  • Reuse OMEN optimizer cache per worker
  • Prefer []byte guess building to cut string churn
• Same generation logic; parallel stdout interleaving may differ across runs
• Rename trainer to pcfg_trainer

[cyclone]

pcfg-go v0.6.0-dev (upcoming release)
https://github.com/cyclone-github/pcfg-go/commit/8fb59c2

While I was testing hashcat's new -a 4 PCFG mode with different hash lists and Rules trained with pcfg-go, @atom mentioned something to me that I couldn't shake: "that needs a fast trainer".
But, pcfg-go was already fast, right? How much faster did it need to be? I was determined to find the answer to this question and pored over possible optimizations in pcfg-go, optimizing keyboard walks, buffers, worker scheduling, profiling hot paths and reducing GC pressure, testing / benchmarking / and comparing Rule outputs to make sure the optimized trainer still produced 1:1 byte-level-identical Rules to v0.5.3.

The result? Let's first back up a minute to run a few comparisons.

• The original python3 pcfg_cracker trainer.py by @lakiw took 13m21s to train rockyou on my Debian 13 benchmarks
• pcfg-go v0.5.3 brought rockyou training down to 2m38s
• And finally, pcfg-go v0.6.0-dev brings rockyou training down to 14.8s

Yes, from 13m21s to 14.8s. Now, that's a fast trainer.

A big thanks to @atom for his support, suggestions, and kindness, and @lakiw for giving the world pcfg_cracker in the first place.