AgentShield: Security scanning for tools behind your LLM proxy

[elliotllliu]

LiteLLM proxies requests to 100+ LLMs, and downstream agents use tools that can have security vulnerabilities.

AgentShield scans AI agent tools, MCP servers, and plugins for backdoors, data exfiltration, and prompt injection before they run.

• 30 security rules with AST taint tracking
• Cross-file data flow + kill chain detection
• Multi-language prompt injection (8 languages)
• 100% offline, zero install, MIT

Real results: scanned 493 Dify plugins, found 6 real backdoors, 0 false positives.

npx @elliotllliu/agent-shield scan ./tool/ --json

Could be useful as a security layer for tool-calling workflows routed through LiteLLM.

https://github.com/elliotllliu/agent-shield