# Agent Audit

A static security scanner for AI agent code that checks for prompt injection paths, unsafe tool inputs reaching subprocess or eval calls, and MCP configuration mistakes, mapped to the OWASP Agentic Top 10.

Agent Audit analyzes agent code built with frameworks such as LangChain, CrewAI, and AutoGen for security issues before it reaches production. It performs tool-boundary taint tracking to catch unsafe inputs that could reach subprocess or eval calls, audits MCP server configuration for credential or access-control mistakes, and applies semantic secret detection, with rules mapped to the OWASP Agentic Top 10 categories.

It installs with pip and runs as a CLI: a scan command analyzes a project, a severity flag filters what gets reported, and a fail-on flag sets the severity that should fail a CI run. On a labeled validation set, the project reports roughly 69 percent precision and 84 percent recall, and documents in its own metrics notes why precision fell and recall rose between two recent releases due to file-discovery fixes rather than a detection regression.

It is intended to run before every merge for teams whose agent code can trigger tools, commands, or external systems.

Website: https://headyzhang.github.io/agent-audit/