Frontier AI training needs a revocable safety case linking risk claims to evidence, veto gates, pause controls, and accountable decisions.
An AI kill switch needs named authority, graded evidence, scoped containment, preserved telemetry, and a tested recovery path.