Enforce zero-trust rules for AI agents to prevent hallucinations, unsafe actions, and policy bypasses