AI agents can misuse legitimate access. Learn why action-level security, trusted policy enforcement and verified approvals are critical to stop data leaks