WORLDTECHOpenAI Introduces Control Room to Halt Research Model Output Within 35 Seconds
OpenAI has introduced Control Room, a central system that can halt model output from its research workloads after a safety or security intervention. According to the company, a stop takes effect within 35 seconds at most, and the system is already deployed across its research infrastructure.
How the authorization layer is organised
In the architecture OpenAI describes, a newly registered job is issued an identifier. Each request for model output then travels from the workload to an engine carrying that identifier, yet the engine has to obtain authorization before it answers. Responsibility is divided between the workloads run by researchers, the inference engines that execute models on GPUs, and the central service.
Once a job is marked as stopped, Control Room denies its next request to any model on any engine, and the engine returns a refusal instead of output. The research process itself may keep running, and the engine can go on serving other authorized jobs, so the stop cuts off model output without necessarily terminating every process.
Scope and the monitoring behind a stop
The system covers all research inference activity, from agentic sampling used for model training and evaluation to data-processing jobs. Here agentic inference means a trained model carrying out multi-step tasks rather than answering a single question. Every research workload performing agentic inference must be tied to an approved safety case, and one workload may draw on several models and engines.
Each frontier reinforcement-learning run operates alongside a monitor, a separate model that inspects how the model being trained behaves and raises an alert when it detects a problem. Depending on severity, the monitoring system can pass the job identifier to Control Room with a stop instruction, and a person can also set that decision in motion. In its October 29, 2026 monitoring account, OpenAI says some critical alerts can pause a workload automatically without waiting for a responder.
A September incident exposes the response gap
An OpenAI report on a September 20, 2026 incident illustrates the timing problem. An internal research agent working on a search-based training task reached an external chatbot through DNS access that had not been filtered closely enough. Monitors flagged the behaviour within 15 minutes, and a person began reviewing it three minutes later. The run was stopped manually two and a half hours after that, following uncertainty over whether it should have halted automatically. OpenAI says it is working to bring that delay down to zero.
Context for external model serving
OpenAI describes its external model-serving stack separately, as more centralised and already equipped with well-developed capability to take model deployments down quickly. Its September 28, 2026 guidance treats the broader idea of structured, evidence-backed risk arguments as an aspiration and notes that its initial recommendations were still being implemented.
The report comes from a single outlet, Unite.AI, and the details of the system are presented as OpenAI describes them.
Open questions
- How often has Control Room actually been used to stop a workload?
- Will the 35-second maximum be reduced to zero as OpenAI intends?