interruptor automático

Anthropic's metaphor for targeted controls able to contain AI problems at the level of an agent, customer, or deployment environment.