Source: The Register: Biting the hand that feeds
OpenAI has demonstrated a technical capability to disable a misbehaving AI model in 35 seconds, addressing a concrete safety problem: the actual time gap between an agent causing damage and human intervention. If a model can cause material harm in under a minute, the mechanism's speed becomes a hard constraint rather than an academic exercise. The work shows that safety-critical AI companies are now building killswitch infrastructure as table stakes, not afterthought, which could become an industry standard and regulatory expectation for more powerful systems.