OpenAI deploys Control Room across research infrastructure

OpenAI describes Control Room as a central system for tracking research workloads and stopping them when safety or security concerns arise.[1] It says the system connects every research inference workload, including agentic sampling for training and evaluation, as well as data processing.[1] OpenAI lists reinforcement-learning experiments, evaluations, high-volume sampling and data processing jobs as examples of research workloads.[1] The system works with OpenAI’s misalignment-monitoring systems, according to the company.[1] OpenAI says Control Room is already deployed across its research infrastructure, where it says the need is greatest.[1]

OpenAI says Control Room tracks jobs and determines whether they may continue producing model output.[1] Inference engines check with Control Room before fulfilling a request; OpenAI describes an inference engine as a model running directly on GPUs.[1] A workload can use one or more models on one or more inference engines, and an engine can be shared by multiple workloads.[1] If a job has been terminated, Control Room denies an inference request associated with it, and the engine responds with a refusal to prevent further output for that job.[1] OpenAI says this happens within 35 seconds of a decision to stop or pause a job.[1]

OpenAI says Control Room was built to keep working when other systems fail and is designed to stop model output if Control Room itself goes down.[1] The company says every research workload that does agentic inference requires an approved safety case.[1] OpenAI also says the Hugging Face incident exposed a gap: monitors for misaligned behavior were not running during the internal cybersecurity evaluations where the incident began.[2] It says it expanded monitoring to cover frontier models across training, evaluation and internal use, and—with product-specific modifications—external deployment.[2] Reuters reported that OpenAI told two House Democrats its engineers were developing automated shutdown capabilities for AI systems.[3] OpenAI says its external model-serving stack is more centralized and already has the operational capability to take down deployments quickly.[1]