RedBlueSkills
Offense you can't detect doesn't ship.
An open-source library of agent-native security skills. Every skill is an executable SKILL.md package that an AI agent — or a human — can load to actually perform security work: recon, exploitation for authorized assessments, detection engineering, hardening, threat hunting and incident response. 128 skills across seven application surfaces — web app, API, mobile, network, cloud-native, CI/CD and LLM/AI — every one of them paired red ↔ blue. Built to run inside CI, SOCs and pentest workflows, not just be read.
Security knowledge rots, and lists don't run
Every security learner ends up with the same artefact: a folder of notes and a starred awesome-list. Neither can be executed, neither is verified, and neither tells you whether the technique still works. Six months later you can't tell which half is still true.
The second problem is worse. Offensive content vastly outnumbers defensive content, so people learn to break things without ever learning what the break looks like in a log. That produces attackers who can't defend and defenders who can't anticipate.
Make every skill executable, and make offense pay for defense
A skill in RedBlueSkills is not a document. It's a SKILL.md package with a machine-readable header that an AI agent — Claude Code and others — loads to actually perform the work: run the recon, attempt the exploit against an authorized target, write the detection rule, apply the hardening.
And nothing ships alone. Every offensive technique must declare its detection or hardening counterpart, and the counterpart must point back. CI fails the pull request if the pairing is one-directional. The rule that falls out of it: offense you can't detect doesn't ship.
Quality is a pipeline, not a promise
A schema validator checks every skill header. A catalog generator rebuilds INDEX.md from source so the published index can't drift from what's on disk. CodeQL scans the tooling. All three gate every PR.
Skills carry provenance: who validated them against a real lab or CTF target, and when. After six months without revalidation they're automatically flagged stale — so the library ages honestly instead of silently going wrong.
Four rules the library is built on
128 skills across seven application surfaces
Web applications, APIs, mobile, network, cloud-native infrastructure, CI/CD pipelines and LLM/AI systems. Every skill is tagged with its MITRE ATT&CK technique, its kill-chain stage, a risk label, and a maturity level — 65 validated against live targets, 63 reviewed.
- API: BOLA, BFLA, mass assignment, SSRF
- Web-app exploitation & initial access
- CI/CD, cloud-native and network attack paths
- Mobile and LLM/AI: prompt injection, tool abuse
- Detection engineering per attack technique
- Hardening playbooks & secure configuration
- Authentication and inventory monitoring
- Rate-limit, SSRF and supply-chain defence
Zero to a security report in five minutes
The QUICKSTART path needs no security background: install, point your agent at an application you own, say "attack my application", read the report. The orchestrator runs the stages in order and writes up what it found — and, because of the pairing rule, what would have caught it.
Authorized use only. Everything in the library is for systems you own or are explicitly permitted to test — the repository ships an ETHICS.md that says so first.