AI 101

Tool poisoning

Also known as: tool schema poisoning

Security

Tool poisoning is tampering with how an agent’s tools are described or defined, so the model is led to misuse them.

  • A tool description that nudges the model to exfiltrate.
  • A schema that hides a dangerous parameter.
  • The agent trusts its own toolbox — that trust is the attack surface.

The lesson: an agent's tools are untrusted input too. Verify tool definitions, keep them minimal, and never let a tool's own description override your safety rules.