Shannon Pentest Skill
Qicaoji7/shannon-pentest-skill
Pentest white-box em uma skill Claude com rigor de PoC
Ver no GitHubSkill de penetração web-app white-box para Claude Code que implementa a metodologia de cinco fases do Shannon AI, permitindo auditorias de segurança end-to-end sem dependências externas. Enfatiza proof-of-concept real: só reporta vulnerabilidades ativamente exploradas, descartando falsos positivos.
Instalação
npx convez-ai Qicaoji7/shannon-pentest-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auditoria de segurança de aplicações web próprias
- Identificação de injection, XSS, broken auth/authz e SSRF
- Validação de vulnerabilidades com proof-of-concept automatizado
Veredito

Ferramenta séria de pentesting que traz rigor metodológico (Shannon) para dentro de Claude Code. Útil para quem precisa auditar apps próprias com disciplina profissional, mas exige autorização legal clara e instância não-produção — não é brinquedo.
README
A tradução para o português deste README ainda está na fila.
shannon-pentest-skill
A Claude Code skill that turns white-box web-application penetration testing into something a single Claude session can run end to end — no separate pentest service, no Docker, no extra API key.
It is an unofficial distillation of Keygraph's open-source Shannon AI pentester: its five-phase pipeline, per-vulnerability-class playbooks, proof-of-exploitation discipline, and three-state (EXPLOITED / POTENTIAL / FALSE-POSITIVE) honesty model, compressed into a single SKILL.md that Claude loads on demand. Not affiliated with or endorsed by Keygraph.
The one law
Only report what you can actively exploit with a working proof-of-concept. Code is ground truth for where a flaw is; live behaviour is the final judge of whether it's real. An unproven "vulnerability" is worse than no finding — it manufactures false confidence.
What it does
Given a web app / API you own plus its source, it runs five phases:
- Preflight + auth — confirm reachability; one real login, reused for all probes.
- Recon — build a security-architecture baseline and a sink inventory from the source (every place user input reaches a dangerous operation).
- Per-class analysis — trace source→sink for injection, XSS, broken authentication, broken authorization, and SSRF; decide vulnerable-vs-safe from the code.
- Exploitation — weaponize each candidate against a non-production instance until a real PoC (proof level 3+) confirms it, or you've disproven it.
- Report — confirmed findings only, each with reproducible steps and a closing fix → regression-test → re-verify loop.
Install
mkdir -p ~/.claude/skills/whitebox-pentest
cp SKILL.md ~/.claude/skills/whitebox-pentest/
Then ask Claude Code for a security audit / pentest / "查漏洞" of an app you own — the skill triggers on that intent.
Safety
- Authorized targets only — apps you own or are explicitly authorized to test.
- Non-production only — exploitation mutates data; stand up a local instance and point at that.
- External-attacker viewpoint — report only what's reachable and exploitable from outside.
Provenance & license
The methodology is adapted from Shannon's agent prompts, phase orchestration, OWASP WSTG coverage map, and proof-by-exploitation safety model. Shannon runs as a Dockerized multi-agent system driven by Claude; this skill is a lightweight re-expression of its playbook so one Claude session can apply the same discipline, without running Shannon itself.
Shannon is licensed under AGPL-3.0. In keeping with that license and as a courtesy to the original authors, this distillation is released under the same terms — see LICENSE and NOTICE. Full credit to Keygraph for the original work.