Offline · stdlib-only · MIT
Guides
Practical writeups on scanning AI agents for prompt injection, MCP tool drift, secret leaks, and other agentic risks. Offline, standard library only.
Prompt injection through tool results: how to detect it in your agent
Tool results sit next to the system prompt. Scan them before the next turn.
Pinning MCP tools: catching silent tool drift and rug-pulls
Treat tools/list like a lockfile. Fail CI when a tool quietly changes.
A 10-minute security audit for your AI agent
No red team required. Six offline scanners, standard library only.
These guides ship with the agentsec toolkit. The scanners are experimental heuristics meant to help a human reviewer. A clean result does not mean something is safe.