Artificial intelligence news.
Straight from the original announcements.

OpenAI introduces Control Room to halt research AI workloads

November 9, 2026, 14:00By the AI News Bites Editorial TeamSource: OpenAI

OpenAI has introduced Control Room, a central system designed to track model workloads across its research infrastructure and stop them when safety or security concerns arise. The company says the system can halt model output within 35 seconds after a decision to intervene.

Key points
  • Control Room tracks OpenAI research inference workloads.
  • It can deny inference requests after a job is terminated.
  • Model output can stop within a maximum of 35 seconds.
  • The system is designed to fail closed if Control Room goes down.
  • Fail-closed behavior is rolling out across remaining research workloads.

A central control layer for research workloads

Control Room connects every research inference workload at OpenAI, including agentic sampling for model training and evaluation, data processing and other jobs. The system works with OpenAI's misalignment monitoring systems and is designed to continue operating when other systems fail.

OpenAI says Control Room is already deployed across the parts of its research infrastructure where the need is greatest. The company's external model-serving stack is more centralized by nature and already has operational capabilities to take down model deployments quickly if needed.

How Control Room can stop model output

The system links three components: the workload, the inference engine and the centralized Control Room service. A workload can be a reinforcement learning experiment, an evaluation, a high-volume sampling job or a data processing task. OpenAI requires every research workload that performs agentic inference to have an approved safety case.

When a job starts, its code requests a job_id from Control Room. Each time the workload asks an inference engine to sample tokens, it passes that job_id to the engine. The engine then asks Control Room for permission to fulfill the request. If the job has been terminated, Control Room denies the request and the engine refuses to produce further model output for that job.

OpenAI places the termination mechanism in the inference layer rather than relying on workload code. The company says research workload code changes frequently and is decentralized, while inference code paths are fewer and more centralized. This design is intended to ensure that research inference requests are subject to Control Room authorization checks.

Monitoring, fail-closed behavior and the 35-second limit

OpenAI says all of its frontier reinforcement learning jobs run with a monitor, which is another model that inspects the model being trained and raises an alert when it detects a potential problem. If the monitoring system decides that a job should stop, it sends the job_id to Control Room. The next inference request from that job, regardless of which model or engine it uses, is denied.

The maximum time between Control Room receiving an instruction to stop a job and further progress becoming impossible is 35 seconds, according to OpenAI. The company attributes the delay to a caching layer used to support the required traffic volume and says it is working to reduce the time to zero.

Control Room also uses a fail-closed setup. If the service goes down, or is disabled through system failure, tampering or another reason, inference requests are denied until it resumes. OpenAI says fail-closed behavior is now rolled out to all frontier model research workloads and will finish rolling out across the remaining research workloads over the next few days. The company plans to strengthen and expand the system alongside further investments in monitoring for misaligned behavior and security breaches.

Related news

Check Point Says AI Agents Need Steering Before the Kill Switch
Safety and ethicsNovember 19, 2026, 14:00

Check Point Says AI Agents Need Steering Before the Kill Switch

Check Point is positioning agentic steering as the first response to risky AI agent behavior, while reserving automatic shutdown for agents that have been hijacked. The…

Princeton study finds ChatGPT’s election answers vary by user
ResearchNovember 18, 2026, 17:09

Princeton study finds ChatGPT’s election answers vary by user

ChatGPT’s memory settings can change the partisan slant of its election-related answers, according to a preliminary Princeton field audit of 160 US users. The study…

Safety and ethics
Safety and ethicsNovember 9, 2026, 14:00

OpenAI deploys Control Room to stop research model workloads

OpenAI has deployed Control Room, a central system that tracks model workloads across its research infrastructure and can stop them when safety or security concerns…