A family of proposals before the US Congress that would require developers of frontier AI systems above a defined threshold to maintain a capability to halt those systems, with substantial penalties for non-compliance. Thresholds have variously been expressed in terms of developer revenue, training compute, or a combination.
None of these measures has been enacted. They are tracked here because the design questions they raise — what specifically must be capable of being stopped, who may order it, and how compliance is tested — are already shaping how frontier developers describe their own controls.
Status should be confirmed against the congressional record before being relied on; bills of this kind are frequently reintroduced in modified form across sessions.