Definition
A security control that sits directly in an AI agent's execution path and can pause or block a risky action the instant the agent tries to take it — like a real-time trip switch, rather than reviewing logs afterward. It's distinct from permission systems that decide access in advance.
Why it matters
As AI agents act faster than humans can review, having a control that can stop a bad action mid-flight — not just flag it later — is critical to preventing real damage.