Control Room: A central system to track and stop model workloads across our research infrastructure

OpenAI has developed Control Room, a centralized system designed to track and manage model workloads across its research infrastructure. The system connects every research inference workload, including agentic sampling for training and evaluation, as well as data processing. Its primary function is to detect potential safety or security issues and halt model output within 35 seconds of an intervention decision. Control Room is engineered to remain operational even during system failures, ensuring model output can be stopped if the system itself goes down.

Control Room integrates three key components: the workload, the inference engine, and the centralized Control Room service. The workload represents jobs such as reinforcement learning experiments or evaluations, while the inference engine functions as the model's operational core, running on GPUs. Each research workload involving agentic inference must be linked to an approved safety case. The Control Room service monitors these jobs and controls their ability to generate model output by verifying job authorization before each inference request.

When a researcher initiates a job, the workload code requests a unique job identifier from Control Room, associating the job with an approved safety case. The workload then attaches this identifier to its inference requests. Before fulfilling any request, the inference engine queries Control Room to confirm the job's authorization status. If authorized, Control Room grants permission, allowing the engine to process the request and generate a response.

If a job is terminated, Control Room marks its identifier as stopped. Subsequent inference requests from the workload are denied, forcing the engine to refuse output. This ensures halted jobs cannot produce further model output while other authorized jobs continue unaffected. Monitors, such as those used in frontier reinforcement learning jobs, inspect model behavior and raise alerts for detected issues, as demonstrated in a recent DNS-related incident.