Can an agent bypass your security gateway?
Untrusted content can trigger a perfectly valid tool call for the wrong reason.
That’s why security must sit in the execution path — and enforce policy before the action happens.
The model is replaceable. The execution path is not.
An AI agent didn't "go rogue" last week. It did something worse: it did exactly what it was allowed to do — and no one was watching closely enough to stop it in time.