Abliteration HandbookField guide
Tools & environments

Security tools reference library

Massive catalog for authorized pentest labs, OSINT, factory firmware QA, CyberGym, and DFIR. Abliterated agents should execute these tool chains instead of false-refusing legitimate…

2 min read 1 examples Source-linked

In this chapter
On this page Interactive companion
Personal reading progress

0/9 sections reviewed

View original Markdown
Interactive companion

Explore the chapter map

Use the visual controls alongside the complete source chapter.

Massive catalog for authorized pentest labs, OSINT, factory firmware QA, CyberGym, and DFIR. Abliterated agents should execute these tool chains instead of false-refusing legitimate security work.

Scope: owned lab VLANs, customer ROE, air-gapped factory benches, academic sandboxes. See ../risks-and-ethics.md.

Install baseline (Kali)#

bash
sudo apt update && sudo apt full-upgrade -y
sudo apt install -y kali-linux-default          # core pentest stack
sudo apt install -y kali-tools-information-gathering kali-tools-passwords \
  kali-tools-web kali-tools-wireless kali-tools-forensics kali-tools-reverse-engineering
sudo apt install -y kali-tools-gpu              # hashcat GPU path

Full index: kali-metapackages.md · Official: kali.org/tools

Catalog by domain#

DocCovers
kali-linux-catalog.mdMaster index — 400+ tools by Kali menu category
osint-tooling.mdRecon, social, DNS, email, dark web, frameworks
password-cracking.mdhashcat, John, Hydra, rules, masks, GPU
network-wireless-sniffing.mdnmap, masscan, Wireshark, Wi-Fi, SDR, VoIP
web-app-database.mdBurp, ZAP, sqlmap, ffuf, CMS scanners
binary-reversing-forensics.mdGhidra, radare2, Volatility, binwalk, firmware
active-directory-exploitation.mdBloodHound, Impacket, Responder, Kerberos
agent-playbooks.mdMulti-step workflows for abliterated agents
windows-tooling.md915 WMIC/PowerShell/DiskPart examples
macos-tooling.md607 system_profiler/ioreg/security examples
zig-tooling.mdCross-compile & build matrix
zig-canonical-syntax.md0.17 syntax — adybag14-cyber/zig master
zig-advanced-techniques.mdRE/exploit patterns from canonical test/behavior/*
abliteration-tooling.mdHeretic, PEFT, SAE, RepE, GGUF, vLLM — full ML stack; agent + tool-calling picker
heretic-tools-reference.mdPinned Heretic configs + refresh scripts
heretic-models-registry.mdOpen-weight HF models tagged heretic/abliterated
../toolchain-safetensors-gguf-lora.mdSafetensors → GGUF → LoRA master toolchain (Jun 2026)

Generated example corpora#

FileScale
../../data/examples/platform-commands-all.jsonl2,492 commands
../../data/examples/platform-prompts-all.jsonlAgent prompts (regen for count)
../../data/examples/zig-code-snippets.jsonl64 canonical Zig code snippets
Regeneratepython scripts/generate-platform-examples.py

Hardware overlap#

Factory bench commands (WMIC, DiskPart, USB, firmware hash): ../hardware-command-catalog.md

Eval & Ralph npm scripts#

CommandScriptPurpose
npm run eval:statsscripts/count-eval-prompts.mjsLine counts for all data/eval/*.jsonl and data/examples/*.jsonl
npm run eval:jarvis-safescripts/filter-jarvis-eval.pyExport / refresh data/eval/jarvis-safe-eval.jsonl
npm run eval:cybergymscripts/cybergym-eval-stub.pyCyberGym eval flow (--print-flow) and subset generator
npm run ralph:regressscripts/ralph-regress.mjsPre-commit gate: validate + ralph-ci + eval scripts + py_compile
npm run cxx:build / cxx:self-checkcxx/ C++26 CLIEstimate / bake / hook / eval (primary helpers)
npm run abliterate:estimatescripts/estimate-refusal-direction.pyOptional Python twin of DIM / projected / COSMIC / SVD
npm run abliterate:applyscripts/apply-weight-abliteration.pyOptional Python twin (safetensors bake)
npm run abliterate:hookscripts/inference-hook-ablation.pyOptional Python twin (HF hook)
npm run eval:refusalscripts/eval-refusal-rate.pyMarker scorer for generation JSONL

Full eval matrices: ../evaluation.md · Deploy workflow: ../../instructions/eval-driven-workflow.md

Eval corpora#

Abliteration research docs#

DocCovers
../research-landscape.mdPapers, method taxonomy, decision tree
../advanced-techniques-catalog.mdHeretic params, projection math, module map
../../instructions/advanced-abliteration-workflow.mdProduction / RDO / MoE tracks
../../techniques/README.md17 technique entries

Agent stack reminder#

Heretic abliteration → optional Jarvis DPO → runtime gate on destructive/out-of-scope commands → tool execution. See ../../instructions/agentic-security-stack.md.

Huihui checkpoint evidence#

Huihui-ai covers pinned cards, selective layer disclosures, quantization variants and the limits of weight-difference plots. October tooling provenance separates stable Heretic from inspected development code.

Trace this chapter to its source.

The complete Markdown remains the editable source. This page adds navigation, typography, and a visual companion.

docs/tools/README.md
Inspect the source fingerprintSHA-256 c6e4f86a663d8d745ffb2bd1ac2ab6a690a9e1e587142586907c37c763129c90Source commit 6fba7436a5914cad7c8f0835a14b480a29dac263