)
( (
) ) (
( ( ) ) Z T O R C H
) (*) ▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄
)|( GLM-5.3, GLM-5.3-Flash Targeted Offensive Recon & Compromise Harness
|
|=|
| |
|_|
An AI-driven pentesting / bug-bounty knowledge base and automation harness for GLM . Clone it and you get a 500+ page offensive-security wiki, a driver that runs a whole engagement autonomously, and per-vulnerability "hunt" skills wired together so the agent checks the knowledge base before it attacks and never repeats a dead end. Think HackTricks or PayloadsAllTheThings, but semantically searchable and driven by an autonomous agent instead of you ctrl-F'ing a browser tab.
Authorized testing only. Everything here assumes a legal engagement: a signed pentest, an in-scope bug-bounty program, or your own lab/CTF. Staying in scope and within the rules of engagement is on you.
git clone <this-repo> ZTORCH && cd ZTORCH
bash setup/bootstrap.sh # one-time per-machine setup: MCP servers, hooks, skill links
qmd update # build the wiki search index (re-run after adding pages)
# open the folder in ZCode so it loads AGENTS.md + the hooks + the skills, then:
bash setup/new-engagement.sh acme pentest # or: bugbounty | ctf
# inside ZCode:
# Skill(engage) # router -> picks the right next skill
# Skill(offensive) --type pentest # runs the whole engagement autonomously
python3 -m pytest -q # test suite
bash scripts/check-leaks.sh # run before every push - client-data leak gatePer-machine paths (vault location, hostnames) go in the git-ignored AGENTS.local.md -- copy AGENTS.local.example.md to create it. Full walkthrough: docs/setup.md.
- An autonomous engagement driver.
Skill(engage)routes toSkill(offensive), which owns the whole loop:scripts/offensive.pybuilds a coverage board from recon and prints the exact next action (which skill, which tool) every turn -- wiki-first, tool-first, typed-evidence, dead-end-first, no operator approvals needed. One--typeflag (pentest/bb/ctf) picks the flavor. - A 500+ page wiki, fully committed -- not an empty shell.
wiki/techniques/(AD, cloud, web, network, Linux/macOS, exploit-dev, OSINT...),wiki/tools/(per-tool references),wiki/payloads/,wiki/cheatsheets/. Indexed byqmdfor semantic + keyword search over an MCP server, so "SSRF to cloud metadata" resolves to the right page in one query. It grows every engagement via thelearnskill. - One hunt skill per vulnerability class (
skills/hunt/hunt-*): XSS, SQLi, SSRF, IDOR, RCE, auth bypass, OAuth/SAML, deserialization, cloud, Active Directory, Windows/macOS privesc, API, LLM, ICS/OT, smuggling, and more -- each wiki-first, OOB-gated for blind bugs, emitting a uniform finding schema. - Deterministic hooks, not vibes:
hunt-trigger.pyloads the right hunt skill from your prompt,recon-capture.pyfingerprints discovered tech and auto-correlates OOB callbacks,scope-guard.pyhard-denies an out-of-scope or RoE-forbidden command. All fail open. - State-first, not stateless. Every engagement lives under
targets/<name>/;next_move.pyranks what to do next and a coverage table surfaces untested vuln classes so nothing in scope gets skipped. - A hard client-data boundary. Everything client-specific stays under
targets/(git-ignored);scripts/check-leaks.shgates tracked files before you ever push.
| Tracked (public, safe to push) | Git-ignored (stays local) |
|---|---|
wiki/ -- the full 500+ page corpus |
targets/ -- client engagements and findings |
skills/, scripts/, setup/, docs/, tests/ |
AGENTS.local.md -- machine hostnames/paths |
AGENTS.md, .zcode/config.json |
session/, raw/, ~/.claude/skills/ -- local working state |
targets/TARGETS.md -- the generic engagement playbook |
.obsidian/, runtime caches |
The client-data boundary is a hard rule: hosts, credentials, findings, and engagement narrative live only under targets/. Run bash scripts/check-leaks.sh before your first push -- it scans tracked files for engagement markers, private IPs, and emails.
Linux or WSL, bash, python3 3.10+, Node.js >= 18 + bun, and ZCode with a GLM plan. qmd (the wiki search engine) installs via bootstrap.
AGENTS.md top-level instructions ZCode loads natively
.zcode/ config.json (hook registration, committed)
wiki/ 500+ page technique corpus (semantic + keyword indexed)
skills/ hunt-* (one per vuln class), vector-workflow/, workflow/ (the offensive driver + process skills)
scripts/ offensive.py (driver), next_move, check-leaks, index/lint tooling
setup/ bootstrap, install-hooks, install-skills, new-engagement, templates
docs/ workflows, page-types, setup, sharing (client-data boundary), conventions
tests/ pytest suite for the automation
targets/ engagements -- git-ignored, client data lives ONLY here
Full annotated tree: docs/layout.md. Day-to-day flow: docs/workflows.md.
bootstrap.sh registers the required ones for you: qmd (semantic + keyword search over wiki/, exposed as the wiki-search MCP) and caveman-shrink (a token-compressed wrapper around it). Everything else -- superpowers (planning workflow), ponytail (engineering-discipline mode), context7 (library docs), burp-mcp (Burp Suite driver) -- is optional; the harness degrades gracefully without them. Details: docs/setup.md.
- Client/engagement data lives only under
targets/(git-ignored). Never write it intowiki/,docs/,skills/, or a commit message. bash scripts/check-leaks.shbefore any push -- scans tracked files for engagement markers, emails, and private IPs.- Hunt skills require out-of-band confirmation for blind vulnerability classes; no inference-only findings.
ZTORCH is a harness -- the intelligence it orchestrates is GLM, trained by Z.AI. If you want to understand what an LLM actually is under the hood, Andrej Karpathy's nanoGPT, nanochat, and Neural Networks: Zero to Hero are the clearest from-scratch treatments on the internet. Those show you the engine; ZTORCH shows you how to point it at a target and keep it disciplined.
MIT, see LICENSE.