Definition
A built-in way to immediately stop a frontier AI model from operating if something goes wrong, including mid-training or during internal use. Several governments and leading labs have committed to implementing these as a concrete condition of operating frontier models. It is the practical 'big red stop button' for catastrophic-risk concerns.
Why it matters
A reliable shutdown is the backstop for responding to an anomalous or dangerous model, and it is moving from a talking point to a signed commitment.