AI agents don't have to break security controls to go rogue. Here's how to keep them in check

Isolate reusable credentials from the reasoning layer. Never pass long-lived secrets or refresh tokens directly to an LLM.