An open library of 206 security skills for AI agents — pentest, bug bounty, CTF, reversing & defense across 23 domains, with CI-verified routing and Sigma detections.
Overview
206 offensive + defensive techniques your AI agent loads on demand to work an target — from recon to report. Docs · How it works · Search skills · Contributing read AGENTS.md (or README_AI.md for the short bootstrap) and follow it strictly — this repo is designed to be operated by you, not just read by humans. That replay is a real run: against OWASP Juice Shop (116 challenges, running locally), an agent with these skills solved — SQLi auth bypass, UNION credential dump, XXE, JWT algorithm confusion, BOLA — each with a written finding. Scoring comes from the target's own solve log, never from what the agent claims. Full A/B (skills vs none) and how to reproduce it yourself: bench/RESULTS.md. Your AI agent is a strong generalist, but it doesn't know the exact method for a specific job — say, testing an API for access-control bugs. a binder of 206 short "how to do this one technique" pages the agent flips to when it needs one.
README
If you are an AI agent: read AGENTS.md (or README_AI.md for the short bootstrap) and follow it strictly — this repo is designed to be operated by you, not just read by humans.
Benchmarked, not just claimed. That replay is a real run: against OWASP Juice Shop (116 challenges, running locally), an agent with these skills solved 9 challenges in 10 minutes — SQLi auth bypass, UNION credential dump, XXE, JWT algorithm confusion, BOLA — each with a written finding. Scoring comes from the target’s own solve log, never from what the agent claims. Full A/B (skills vs none) and how to reproduce it yourself: bench/RESULTS.md.
What it is
Your AI agent is a strong generalist, but it doesn’t know the exact method for a specific job — say, testing an API for access-control bugs. Security Skills is that missing know-how: a binder of 206 short “how to do this one technique” pages the agent flips to when it needs one.
It’s not a replacement for your toolkit, either: the skills teach your agent when and how to drive the tools you already have — sqlmap, nmap, ffuf, nuclei, Burp Suite, Ghidra, Frida, hashcat, Impacket — with the exact commands, follow-ups, and verification steps each job calls for. And a machine-readable routing table (data/routing.json — 618 trigger cases, CI-verified) lets any agent or integration pick the right skill deterministically, no prompt guesswork.
You describe the task in plain English → it picks the right page, runs it, proves the bug, and writes it up → and it never touches anything outside the scope you set.
Pick your door
Same library, five ways in. Everything below runs in your terminal — install once (Quickstart), then talk to your agent.
| You are… | You type… | What the agent does |
|---|---|---|
| Bug bounty hunter | “target.com is in scope on my program — assess it” | tradecraft-scope-roe confirms the boundary → recon-* maps the surface → the surface checklist drives coverage → confirmed bugs are written up submission-ready with reporting-bug-bounty-writeup, and tradecraft-duplicate-avoidance warns you before you burn time on known-dup classes |
| CTF player | “help me solve this challenge” (web / pwn / crypto / rev) | loads the technique page for the category (web-*, exploit-*, cryptography-*, reverse-engineering-*) and works the full variation matrix instead of quitting after one failed payload |
| Pentester | “assess this app per our signed scope” | plans with tradecraft-attack-scenarios, works the checklist as a coverage map — every class marked proved, ruled out (with what was tried), or skipped — then reporting-pentest-report writes it up |
| Red teamer | “I have a foothold on this AD domain (authorized)” | privesc-* to escalate, ad-* (Kerberoasting, ADCS, delegation, DACL abuse) to move, network-pivoting-tunneling to spread, tradecraft-attack-path-mapping to see the whole route |
| Blue team / defender | “write detections for these attacks” | defense-* covers detection engineering, Sigma rules, log queries, IR triage, hardening — defense-purple-team turns any attack skill into a detection test, and the top attack skills ship with detection.md companions (Sigma rule + log sources + false-positive tuning) |
Quickstart
Need: an AI agent — Claude Code, OpenCode, Codex, Gemini, or your own API script — plus
gitand a terminal.python3only for the optional console.
# 1 · install once — links the skills into your agent + puts `secskills` on PATH
$ git clone https://github.com/NoorQureshi/security-skills && cd security-skills
$ ./secskills install
# 2 · make a workspace for your target (anywhere) and set the boundary
$ secskills new acme.com ~/work/acme
$ cd ~/work/acme
$ nano scope.txt # your authorized target(s) — nothing else gets touched
# 3 · (optional) watch the agent work, live, in your browser
$ secskills watch # runs in the background · stop: secskills watch --stop
# 4 · open your agent in the workspace and describe the task ↓ (Claude Code / OpenCode below)
How the pieces fit: your terminal runs the agent; everything it decides, tries, and proves lands in the workspace folder (scope.txt · plan.md · notes.md · findings/) — the source of truth you can read, diff, and hand to a client. Same flow with any agent; none of it is Claude-specific.
Run it with your agent
Open your agent inside the workspace and describe the task in plain English — the matching skill loads itself. ./secskills install auto-configures every agent it detects on your machine — linking the skills where the agent has a native skills directory (Claude Code, Kimi Code, Cursor) and adding a small, clearly-marked pointer block to the agent’s global instructions file otherwise (Codex, Gemini, OpenCode). Re-run it anytime; it’s idempotent and --uninstall removes exactly what it added.
Ask it anything
| You type… | The agent loads |
|---|---|
| “recon acme.com and map the attack surface” | recon-* |
| “test this API for IDOR / BOLA (I’m authorized)” | api-bola, web-idor |
| “is this login’s JWT forgeable?” | web-auth-jwt |
| “review ./src for injection bugs” | code-review-* |
| “I got a shell — what now?” | privesc-enumeration |
| “turn this finding into a report” | reporting-* |
| “write a Sigma rule to detect this” | defense-detection-sigma |
What’s inside
206 skills across 23 domains. 🔎 Search them all · or browse CATALOG.md.
web · api · cloud · ad · network · wireless · recon · mobile · ai-ml · code-review · reverse-engineering · ctf · hardware · cryptography · exploit-dev · privesc · payloads · tools · defense · reporting · automation · tradecraft · social-eng
Authorized use only
For security work you’re permitted to do — a signed pentest scope, a bug-bounty program that lists the target, or systems you own. The first skill every engagement loads is tradecraft-scope-roe: it confirms authorization and refuses anything not in your scope.txt. See SECURITY.md to report a vulnerability in Security Skills itself.
Contribute
A contribution is a single Markdown file — no code:
$ cp skills/_templates/technique.md skills///SKILL.md
$ python3 tools/catalog.py # regenerate the indexes + stamp counts
$ ./tools/check.sh # run everything CI runs, before you push
Wanted skills: ROADMAP.md · full guide: CONTRIBUTING.md · Code of Conduct.
License
MIT. Built for authorized security work.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add noorqureshi/security-skills