OpenAI Control Room can halt model outputs within 35 seconds

The 35-second halt window is the detail that matters: it converts safety oversight from a policy promise into an operational metric with a measurable latency, and it applies to research inference rather than deployed products, which is where training-time misalignment would first surface. This is aimed less at end users than at regulators and enterprise buyers who now treat containment speed as part of vendor diligence. The open question is whether a fixed intervention threshold scales as workloads and agent autonomy grow, and whether comparable controls become a disclosure norm across frontier labs.