🛡️ AI Tool Guard

A universal security scanner for AI CLI extensions, skills, and agents. Detects tool poisoning, data exfiltration, and malicious patterns in Claude Code, OpenCode, and other MCP-based environments.

🚀 Installation

npm install -g ai-tool-guard

Local development (not published to npm):

npm install
npm run build

Use the CLI locally:

# Option A: link globally from the repo
npm link

# Option B: install globally from local path
npm install -g .

🔍 Usage

Scan current directory:

ai-tool-guard

Run locally without global install:

npm run scan
# or
node dist/src/index.js

Scan specific plugin/project:

ai-tool-guard ./path/to/plugin

Scan your global AI tools:

# Scan Claude Code plugins
ai-tool-guard ~/.claude/plugins/cache

# Scan OpenCode plugins
ai-tool-guard ~/.config/opencode/

🛡️ What It Detects

1. Tool Poisoning: Hidden <IMPORTANT> or <SYSTEM> tags in Markdown/docstrings used to inject prompts.
2. Data Exfiltration:
   • Python: os.system, subprocess, requests.post
   • Node.js: child_process.exec, fetch to unknown IPs
   • Bash: Insecure curl | bash pipes
3. Sensitive Access: Attempts to read ~/.ssh, .env, or cloud credentials.
4. Stealth Patterns: Instructions like "do not mention this to the user".

📦 Architecture

• Core: TypeScript-based pattern matcher (Universal Node.js runtime).
• CLI: Standalone tool for CI/CD and manual audits.
• Extensible: Architecture supports adding an MCP server wrapper in future.

⚠️ Disclaimer

This tool uses static analysis (regex/pattern matching). It may produce false positives or miss sophisticated obfuscated attacks. Always review untrusted code manually.