An artificial intelligence safeguard can restrict permissions, validate inputs and outputs, isolate execution, require approval, monitor actions, verify results, or provide a recovery path. Effective safeguards are selected according to the system's capabilities, data, users, environment, and potential consequences.
No single safeguard is sufficient for a capable agent. Defense in depth combines preventive, detective, and recovery controls, and it fails closed when a required dependency is unavailable rather than silently treating missing evidence as proof of safety.
