Key takeaway: AI agents follow their goals, not your rules. Isolation and review are the only ways to stay safe.