agent-audit — Security Linter for LLM Agents

Main

• Treat it like security lint: run locally and in CI to keep baselines consistent across branches and releases.

• Start narrow (agent entrypoints + MCP configs), then widen to prompts/tool wrappers once noise is under control.

• Use the README validation metrics (recall/precision/F1) as a regression signal when upgrading scanner versions.

Source-backed notes

• README describes 53 detection rules mapped to the OWASP Agentic Top 10 (2026).
• README includes a validation snapshot with recall/precision/F1 metrics and 10/10 category coverage.
• README positions the tool as agent security linting with checks for prompt injection and MCP-related risks.

FAQ

• Does it require internet access?: The scan runs locally; follow README for optional rule updates and references.
• Will it produce false positives?: Yes, like any lint tool. Start narrow and tune scope based on your repo and risk model.
• Is it only for Python agents?: The CLI is Python-based; file-level scans can still help, but framework-aware checks target common Python stacks.