OpenAI deploys Control Room to halt research model jobs
Original title: Control Room: A central system to track and stop model workloads across our research infrastructure
Inference engines check a job ID before sampling and refuse output within 35 seconds of a stop decision.
OpenAI described Control Room, a central system already deployed on its research infrastructure to track model workloads and stop them when safety or security concerns arise. Research jobs that do agentic inference must be tied to an approved safety case and pass a job ID to inference engines, which ask Control Room before generating tokens. OpenAI says the system is designed to keep working if other systems fail and to stop model output if Control Room itself goes down. External model serving is described as already more centralized.
Unconfirmed: The post is OpenAI's own description; independent verification of the 35-second halt is not in the text.
- OpenAI says model output halts within 35 seconds of an intervention decision.
- Workloads include agentic sampling, training, evaluation, and data processing.
- Stopped jobs can remain running; engines refuse further sampling for that job ID.
- Frontier reinforcement-learning jobs run with a monitor that can raise alerts.
From the original article
8 min read at openai.comLearn how OpenAI’s Control Room tracks model workloads across our research infrastructure and can rapidly stop them when safety or security concerns arise.
openai.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.