All updates

• OpenAI deploys Control Room to track and halt research model workloads within 35 seconds

  OpenAI said on 9 November 2026 that it built Control Room, a central system that tracks model workloads across its research infrastructure and can stop them when safety or security concerns arise. The system connects research inference jobs, including agentic sampling for training and evaluation, and is designed to halt model output within 35 seconds of a decision to intervene. Inference engines check Control Room before fulfilling a request, and the integration fails closed so that if Control Room itself goes down, further inference is denied. OpenAI said Control Room is already deployed across research infrastructure, works with misalignment monitors, and that fail-closed coverage is live for all frontier research workloads, with remaining research jobs due to finish rolling out over the following days.

  OfficialNov 9Source: OpenAIPart of Control Room

RSS for this view → · By the numbers