Stopping or correcting an AI system's behavior in real-time during execution, rather than after the fact.