convez.ai
SkillSegurançaIntermediário
TheGrowthShortcut-AI-Builders

Auditoria de Segurança de Repositórios GitHub

TheGrowthShortcut-AI-Builders/repo-security-review

Audita repositórios GitHub antes de instalar — detecta código malicioso automaticamente

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que audita repositórios GitHub em busca de código malicioso, exfiltração de dados ou comportamento suspeito antes da instalação. Funciona automaticamente em Claude Code, Codex, OpenCode e 30+ ferramentas de IA via Skillfish, analisando scripts de instalação, código-fonte e dependências para retornar um veredito claro (Seguro / Cuidado / Inseguro).

Instalação

$npx convez-ai TheGrowthShortcut-AI-Builders/repo-security-review

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Validar segurança de ferramentas open-source antes de usar
  • Detectar tentativas de exfiltração de dados ou credenciais em scripts de instalação
  • Identificar dependências typosquatted ou pacotes com hooks suspeitos de pós-instalação

Veredito

Ferramenta essencial para quem trabalha com código open-source em ambientes com agentes IA. Executa análise profunda e automatizada de segurança, cobrindo vetores comuns de ataque (scripts maliciosos, obfuscação, typosquatting). Integração elegante via Skillfish com suporte a múltiplos agentes.

README

A tradução para o português deste README ainda está na fila.

repo-security-review

A skill that audits a GitHub repository for malicious code, data exfiltration, or sketchy behavior before you install it.

Paste any GitHub URL into your AI coding agent and ask "is this safe to install?" — the skill walks through the repo's install scripts, source, and dependencies and returns a clear verdict (Safe / Caution / Unsafe) with a concrete recommendation.

Works in Claude Code, Codex, OpenCode, Cursor, and 30+ other AI coding agents.


Install

Recommended (one command, works everywhere)

npx skillfish add TheGrowthShortcut-AI-Builders/repo-security-review

Skillfish auto-detects which agent you're using and writes the skill to the right location. No flags, no config — just run it.

Manual install (Claude Code)

If you'd rather skip Skillfish:

git clone https://github.com/TheGrowthShortcut-AI-Builders/repo-security-review.git \
  ~/.claude/skills/repo-security-review

For Codex and OpenCode the install path differs by version — Skillfish handles that automatically, so the one-command path above is the safer route for those.


How to use it

The skill triggers automatically — you don't need to ask "is this safe?" Just hand your agent a repo to download, clone, install, or run, and the audit runs first. It also still fires on explicit safety questions. Examples:

You: Download github.com/some-user/some-tool and set it up
You: Clone this and run it for me
You: Install this CLI
You: Is this safe to install? https://github.com/some-user/some-tool
You: Audit this code for anything sketchy

The agent clones the repo read-only into a temp directory, runs through the audit workflow on the real files, returns a structured verdict, and only proceeds to install/run if it's safe.

Guaranteed auto-trigger (optional hook, Claude Code)

The skill's description makes the agent reach for it proactively, but model-judged triggering can occasionally miss. For deterministic firing on every prompt that mentions a repo URL or a download/clone/install action, install the included UserPromptSubmit hook:

# 1. Copy the hook script
mkdir -p ~/.claude/hooks
cp ~/.claude/skills/repo-security-review/hooks/repo-security-trigger.py ~/.claude/hooks/

# 2. Register it in ~/.claude/settings.json under "hooks":
#   "hooks": {
#     "UserPromptSubmit": [
#       { "hooks": [ { "type": "command",
#         "command": "python3 ~/.claude/hooks/repo-security-trigger.py" } ] }
#     ]
#   }

Restart Claude Code (or run /hooks) to load it. The hook scans each prompt for a github.com/<owner>/<repo> link, a git clone URL, or download/install/setup intent, and injects a reminder forcing the audit to run before anything is downloaded. Unrelated prompts are untouched.


What it checks

Install scripts

  • curl | bash from non-official hosts
  • Remote-code execution (eval $(curl ...))
  • Shell profile injection (.bashrc, .zshrc)
  • Environment variable harvesting ($AWS_SECRET, $GITHUB_TOKEN)
  • Hidden file creation outside install scope

Source code

  • Base64-encoded payloads
  • Reads of ~/.ssh, ~/.aws, browser data, credentials
  • Obfuscated logic, dynamic eval/exec
  • Phone-home / analytics endpoints
  • Hardcoded IPs, dynamic DNS, URL shorteners

Dependencies

  • analytics / tracking / telemetry packages
  • Typosquatted names (lodahs vs lodash)
  • Suspicious postinstall hooks
  • GitHub-URL deps instead of registry deps
  • Recently created or low-download packages

Full pattern catalog: references/red-flags.md.


Sample output

## Security Review Summary: example-tool

### ✅ Install Script - CLEAN
Only copies local files into ~/.config/example. No network calls, no eval,
no shell profile modification.

### ⚠️ Application Code - SUSPICIOUS
Reads ~/.ssh/id_rsa during first-run telemetry init (see src/init.js:42).

### ✅ Dependencies - CLEAN
All deps from npm registry, no typosquats, no postinstall hooks.

### Assessment: CAUTION
The install script is safe but the runtime behavior reads SSH keys without
a documented reason. Recommend: install in a sandbox, or skip until the
maintainer explains src/init.js:42.
Ver o original no GitHub

Bom para

Relacionados