OpenAI deploys Control Room to halt research AI output within 35 seconds

Originally published on The Daily Flare.

OpenAI says it has deployed Control Room across its research infrastructure, a central system that can halt model output within 35 seconds after researchers decide to pause or terminate a job.

The company described the system in a November 9 post, saying it is intended to track research workloads and stop them when safety or security concerns arise. It operates alongside OpenAI’s systems for monitoring potentially misaligned model behaviour, linking detection and intervention to a mechanism that can block further output.

Control Room covers research inference workloads, including agentic sampling used in model training and evaluation, data processing and other jobs. OpenAI requires research workloads that use agentic inference to be tied to an approved safety case. Workloads can include reinforcement-learning experiments, evaluations, high-volume sampling or data-processing tasks.

The system connects three parts of a job: the workload code that runs the experiment, the inference engine that runs a model on graphics processing units, and the central Control Room service. A workload first requests a unique job ID from Control Room. It then attaches that ID whenever it asks an inference engine to generate model output.

Before fulfilling each request, the inference engine checks the job ID with the central service. If Control Room has marked the job as stopped, it denies permission and the engine refuses the request, preventing the model from generating further output for that workload. The same model engine can serve more than one workload, while a single workload may use multiple models or engines.

Putting the check outside the experiment’s own code is a central design choice. Research code is managed by researchers and can change across different kinds of work, OpenAI’s explanation says. A separate enforcement point means a shutdown does not depend on the workload’s code carrying out the instruction. The inference engine applies the restriction when the job next requests model output.

OpenAI says the system is built to keep working when other systems fail, and to stop output if Control Room itself becomes unavailable. In that situation, inference requests are denied until the service is operating again. The company also says it runs continuous automated checks, including sweeps using AI agents, to find inference engines that may have been started without Control Room enabled.

Shutdown delay remains

Control Room does not stop output instantly. OpenAI says it can take up to 35 seconds from a decision to intervene before the model can no longer make progress. Firstpost reported that the delay is tied to cached permissions used to manage inference authorisation. Rather than checking every request against the central service in real time, the infrastructure uses cached permissions to handle high volumes of activity.

OpenAI says the system halts output within that 35-second window and that it wants to reduce the delay further. The company’s stated objective is to make the shutdown process as close to immediate as possible. Until then, the 35-second maximum is the interval the system allows between an intervention decision and the end of model output.

The rollout applies to OpenAI’s research infrastructure, where the company says the need is greatest. OpenAI distinguishes that system from its external model-serving stack, which it describes as more centralised by nature and already supported by operational capabilities to take down model deployments quickly.