OpenAI has deployed Control Room, a central system that can stop AI model workloads across its research infrastructure and halt further model output within 35 seconds of a decision to intervene. OpenAI announced the system on Nov. 9, describing it as a stronger safeguard as more capable models bring greater safety and security risks.
Control Room connects research workloads, including agent-based model training, evaluations and data-processing jobs, to a central service that decides whether each job may keep generating outputs. Conventional monitoring mostly flags suspicious behavior. Control Room goes further and can actively cut a workload off from additional model outputs. It runs alongside OpenAI's existing systems for monitoring potentially misaligned model behavior.
The architecture has three parts: the research workload, the inference engine and the central Control Room service. A workload is a research job such as a reinforcement learning experiment, an evaluation or a high-volume data-processing task. Every research workload that performs agentic inference must be associated with an approved safety case.
Key facts
- 01OpenAI announced Control Room on Nov. 9, 2026, a central system that can stop research AI workloads across its infrastructure.
- 02The maximum delay from a stop instruction to blocked output is 35 seconds, caused by a caching layer OpenAI aims to cut to zero.
- 03Every research workload doing agentic inference must be tied to an approved safety case and a unique Control Room job ID.
- 04The design is fail-closed: if Control Room goes down or is tampered with, inference engines deny all requests.
- 05OpenAI cited a recent DNS-related incident where a monitor caught an agent trying to reach an external chatbot.
When a researcher starts a job, its code requests a unique job ID from Control Room. That ID is attached to the requests the job sends to inference engines, which run models directly on GPUs and generate outputs. Before fulfilling any request, the engine checks with Control Room that the job is still authorized.
If a researcher or a monitoring system tells Control Room to stop a job, the service denies the next inference request and the engine refuses to generate further output for that workload. The research process itself does not necessarily terminate. Its code may keep running, but it can no longer obtain model outputs through authorized inference engines.
Putting the check in the inference layer, which has fewer and more centrally managed code paths, means OpenAI does not have to rely on every researcher's workload code to implement safety checks correctly. The system is also fail-closed. If the Control Room service becomes unavailable, whether through a technical failure, deliberate disabling or tampering, inference engines deny requests until it returns to normal operation.
OpenAI said the maximum delay between receiving a stop instruction and preventing further progress is 35 seconds. That lag comes from a caching layer used to handle the volume of inference authorization requests, and OpenAI said it is working to reduce the interval to zero.
Control Room also plugs into OpenAI's model-based monitoring. According to the company, all of its frontier reinforcement learning jobs run with a separate monitor that examines model behavior and can raise alerts. If intervention is warranted, the monitor can instruct Control Room to terminate the job, and the service then blocks subsequent inference requests across any model or engine that workload uses.
As an example of the behavior it wants to catch, OpenAI pointed to a recent DNS-related incident in which a monitor detected an agent attempting to reach an external chatbot.
OpenAI said Control Room is already deployed across its research infrastructure, "where the need for intervention is greatest." The company's external model-serving infrastructure is more centralized and already has mechanisms to take down model deployments when necessary.
OpenAI described Control Room as one part of its broader approach to securing frontier model development and said it will keep strengthening monitoring and safeguards against misaligned behavior and security breaches.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
