OpenAI's Control Room inference kill switch

OpenAI has deployed Control Room, a centralized system that tracks all research inference workloads and can halt model output within 35 seconds of a decision to intervene. The system integrates with every research inference workload, including agentic sampling, model training, evaluation, and data processing, and denies inference requests for terminated jobs. Control Room is designed to fail closed: if the service goes down or is tampered with, all inference requests are automatically denied until it resumes. Every research workload doing agentic inference must be associated with an approved safety case, and continual sweeps check that no rogue inference engines are running outside Control Room oversight. The termination mechanism sits in the inference layer rather than workload code because there are few inference code paths, making it a reliable bottleneck. OpenAI is working to reduce the current 35-second maximum lag, which results from a caching layer for inference authorization, to zero. OpenAI describes Control Room as one part of a broader approach to safe frontier model training and deployment.