Definition
A dedicated control component that decides, in real time, whether a specific tool call an AI agent wants to make is allowed — a security checkpoint for every agent action. It evaluates each call against policy and blocks anything not permitted, and may cryptographically verify where requests come from. Weaknesses in how such gates verify requests or signatures can let unauthorized actions slip through.
Why it matters
This is the enforcement muscle behind 'guardrails'; a broken policy gate means an agent's promised constraints (for example, 'cannot cancel orders') are not actually real.