
Control Room: OpenAI’s Central Stop System for Research Model Workloads

OpenAI has deployed Control Room, a central system for tracking model workloads across its research infrastructure and halting them quickly when safety or security concerns arise. It connects every research inference workload, including agentic sampling used for model training and evaluation, and data processing. Control Room works alongside OpenAI‘s misalignment monitoring systems and can stop further model output within 35 seconds of a decision to intervene. The system is designed to keep operating even when other systems fail, and it stops model output entirely if Control Room itself goes down. It is already live across OpenAI‘s research infrastructure; the external model-serving stack is already more centralized and has its own operational capability to take down deployments quickly.
Control Room integrates three components: the workload, the inference engine, and the centralized Control Room service. The workload is the job being run, such as a reinforcement learning experiment, an evaluation, or a high-volume sampling/data-processing job. OpenAI requires every research workload that performs agentic inference to be associated with an approved safety case. The inference engine is the model itself, running directly on GPUs; a workload may use multiple models or engines, and the same engine may serve multiple workloads. The Control Room service tracks jobs and decides whether they are allowed to keep generating output. Engines check with the service before fulfilling any request, and if a job is stopped the service denies the request, preventing further output.
When a job starts, the workload code requests a new job_id from Control Room. Whenever it needs to sample tokens, the workload calls an inference engine and passes the job_id. The engine then requests permission from Control Room with that job_id. If the job has been terminated, Control Room denies the request and the engine returns a refusal, stopping further model output for that job.
For frontier reinforcement learning jobs, a monitor model inspects everything the model-being-trained is doing and raises an alert if it detects a problem. OpenAI notes that its recent DNS-related incident was caught by a monitor. The monitoring system can decide to stop a job, sending a stop message with the job_id to Control Room. Control Room marks the job as terminated, and the next inference request with that job_id is denied.
The maximum time between Control Room receiving a stop instruction and the moment further progress is impossible is 35 seconds, caused by a caching layer used for inference authorization at the required traffic volume. OpenAI says it is working to drive this down to zero for instant job termination. Because workloads are owned by researchers and change often, they are not a reliable place for critical safety protections. The termination mechanism is therefore implemented in the inference layer, which is centralized and uniform across OpenAI research. This lets the inference layer serve as a bottleneck, giving strong guarantees that every inference request is subject to Control Room authorization checks. OpenAI also runs continual agentic and programmatic sweeps to detect rogue engine implementations that have Control Room disabled.
Control Room fails closed: if the service goes down, is disabled, or is tampered with by a bad actor, all inference requests are denied until it resumes normal operation. Fail-closed is already rolled out to all frontier model research workloads and will finish rolling out across the remaining research workloads in the coming days. The system is part of OpenAI‘s broader approach to safe frontier model training and deployment, alongside monitoring for misaligned behavior and security breaches.

