Hermes Autonomous Audit System
houssem98/hermes-audit
Auditoria autônoma de contratos com PoCs confirmados e aprendizado baseado em evidência
Ver no GitHubAgent multi-especialista para auditoria de segurança em smart contracts, deployável no Claude Code. Processa URLs de contests (Code4rena, CodeHawks, Sherlock) ou repositórios e gera relatórios de classe contest com PoCs validados e evidências concretas. O sistema melhora iterativamente sem alucinar, priorizando precisão sobre volume de achados.
Instalação
npx convez-ai houssem98/hermes-audit1 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 em smart contracts antes de deployment
- Participação em contests de bug bounty (Code4rena, Sherlock, CodeHawks)
- Validação de invariantes e propriedades em Foundry/Echidna
- Geração de relatórios técnicos com exploits comprovados
Veredito

Ferramenta rigorosa e bem-arquitetada para auditoria de contratos inteligentes, com guardrails fortes (gate humano obrigatório, foco em evidência, sem auto-submit). Exige experiência em Solidity e testes de segurança, mas elimina achados falsos através de loop de aprendizado disciplinado.
README
A tradução para o português deste README ainda está na fila.
Hermes Autonomous Audit System
Self-improving multi-agent smart-contract security audit, deployable in Claude Code.
Paste a Code4rena / CodeHawks / Sherlock link (or a repo URL) → a contest-grade, evidence-backed report with passing PoCs, gated on one human review. The system gets better at finding real bugs every run — it never learns to hallucinate. Evidence is the only currency.
Quick start
# in this repo, with Claude Code:
/audit <repo-path-or-contest-url> # full PHASE 0–8 pipeline → report.md + findings.json
/benchmark [prompt-version] # frozen suite gate before promoting any tuned prompt
How it works
INGEST → ENV → ORCHESTRATOR
├── 12 SPECIALISTS (hunt their class, memory-primed) → findings.json[]
├── INVARIANT/FUZZ ENGINEER (Foundry/Echidna/Halmos) → property breaks
├── POC FORGE (loop until passing PoC or killed) → Confirmed exploits
├── TRIAGE (dedup · FP-kill gate · severity)
└── LEARNING ENGINE (TP/FP/MISS memory → retrieval + gated tuning)
│
▼ HUMAN REVIEW GATE (required before any external submission) ▼
Structure
| Path | Purpose |
|---|---|
CLAUDE.md | Three laws + self-improvement invariant + FP-kill gate + severity matrix + finding schema. Inherited by every agent. |
.claude/agents/ | Orchestrator + 12 specialists + poc-forge + triage + learning-engine. |
.claude/commands/ | /audit, /benchmark. |
.claude/memory/ | Append-only labeled stores: tp/ fp/ miss/ prompt-versions/ benchmark/. |
The three laws (non-negotiable)
- EVIDENCE LAW — no finding without an exact anchor + a passing PoC or an airtight state-transition attack path.
- REACHABILITY LAW — prove a non-trusted actor can call the path. Trusted-admin malice = centralization note, not a High.
- HUMILITY LAW — assume the devs are competent. A killed false positive is a win.
Self-improvement invariant
The learning loop only adjusts where agents look, what they prioritize, and how
they phrase findings. It never lowers the evidence bar, auto-promotes unconfirmed
findings, or optimizes for finding-count. Target = precision-at-confirmed. status
flips to Confirmed only via a passing PoC.
Guardrails
- Human gate before any external submission (no AI-slop auto-submit).
- Authorized scope only (contest repos / engaged clients / bounty programs that permit testing).
- Throwaway sandbox, no secrets/keys, fork tests on read-only RPC.
- Pinned tool + compiler versions stamped per run.