Skip to content

README.en.md

Latest commit

 

History

History
123 lines (82 loc) · 4.38 KB

README.en.md

File metadata and controls

123 lines (82 loc) · 4.38 KB

🛡️ Agent Shield

Lightweight open-source AI Agent security scanner — Static Analysis + Runtime Interception

npm License: MIT Tests Rules

Free, offline, zero-config AI Agent security scanner. Quickly check your skills, MCP servers, and plugins for security issues.

npx @elliotllliu/agent-shield scan ./my-skill/

🏆 Key Features

1. 🔒 Runtime MCP Interception

Monitor MCP tool calls in real-time — not just static code scanning.

agent-shield proxy node my-mcp-server.js
agent-shield proxy --enforce python mcp_server.py

Detects: tool description injection, result injection, credential leaks, sensitive path access, behavioral anomalies.

2. ⛓️ Cross-File Attack Chain Detection

Tracks complete attack paths across multiple files.

3. 🆓 Free & Offline

No account · no code upload · no API key required · npx one-liner.

🧠 Context-Aware Scoring

  • SDK Awareness: Auto-detects 25+ SDKs — network calls via known SDKs get lower risk scores
  • Auth Flow Recognition: Identifies OAuth2, JWT, session management — legitimate auth isn't flagged as exfiltration
  • Data Flow Tracking: Traces variables from source to sink — only flags actual exfiltration paths
  • Confidence Scoring: high/medium/low confidence multipliers reduce false positive impact

⚡ Quick Start

npx @elliotllliu/agent-shield scan ./skill/      # Scan
npx @elliotllliu/agent-shield proxy node server.js # Runtime proxy
npx @elliotllliu/agent-shield install-check        # Check installed agents

📊 Positioning

Agent Shield is a lightweight open-source tool for quick security self-checks during development. It is not a replacement for enterprise security platforms.

Agent Shield Snyk Agent Scan Tencent AI-Infra-Guard
Positioning Lightweight OSS tool Commercial service Enterprise red team
Runtime interception ✅ MCP Proxy ❌ ❌
Cross-file attack chain ✅ ❌ Partial
Static rules 31 6 Many (incl. infra)
Offline/free ✅ ❌ Account required ✅ Open source
Zero config ✅ npx one-liner ❌ Python+uv needed ❌ Docker needed
Security team ❌ Individual project ✅ Snyk security team ✅ Tencent labs
Vulnerability DB ❌ ✅ ✅
Enterprise support ❌ ✅ ✅ Pro version
Production validated ❌ Early stage ✅ ✅ Black Hat
VS Code / Action ✅ ❌ ❌

Good for: Dev-stage quick checks · CI/CD lightweight gate · Runtime MCP monitoring Not for: Enterprise compliance as sole tool

🔍 31 Security Rules

🔴 High: data-exfil · backdoor · reverse-shell · crypto-mining · credential-hardcode · obfuscation

🟡 Medium: prompt-injection (55+ patterns) · tool-shadowing · env-leak · network-ssrf · phone-home · toxic-flow · skill-risks · python-security (35 patterns) · go-rust-security (22 patterns)

🟢 Low: privilege · supply-chain · sensitive-read · excessive-perms · mcp-manifest · typosquatting · hidden-files

Advanced: cross-file · attack-chain · multilang-injection · python-ast · description-integrity · mcp-runtime

📈 Testing

Benchmark: 120 self-built samples, 100% F1. ⚠️ Not independently verified.

Dify Plugins: Scanned 493 plugins, found 6 true high-risk (eval()/exec()), 0 false positives at high severity.

Usage

agent-shield scan ./skill/ [--json|--sarif|--html] [--fail-under 70] [--ai]
agent-shield proxy [--enforce] [--rate-limit N] <command>
agent-shield mcp-audit <command>
agent-shield discover

GitHub Action · VS Code Extension

Links

📦 npm · 📖 Rules · 🇨🇳 中文

🌐 Community & Partners

Partner Contribution
Agent Skills Hub Real-world testing across skill registries, security insights, and feature feedback

License

MIT