Control Room: A central system to track and stop model workloads across our research infrastructure | OpenAI

Open original

OpenAI built Control Room, a central system that tracks every model workload across its research infrastructure and can stop one within 35 seconds when a safety or security issue comes up. It fails closed, so if the system goes down, inference requests are denied rather than allowed to continue.

  • Control Room connects all research inference workloads, including agentic sampling for training and evaluation, and works alongside OpenAI's misalignment monitors, which are themselves models that watch what training runs are doing.
  • Termination happens in the inference layer, where nearly all research model calls pass through a small number of code paths, so every request can be checked for authorization. The current lag is 35 seconds, caused by a caching layer, and OpenAI says it aims to bring that to zero.
  • The system fails closed: if Control Room crashes or is disabled, including by tampering, all inference requests are denied until it resumes. OpenAI says this is already live for frontier model research workloads, with the rest rolling out over the following days.

OpenAI's Control Room is a central system that tracks every inference workload running across its research infrastructure and can halt them quickly when safety or security problems show up. It connects agentic sampling used in model training and evaluation, data processing, and similar jobs. The stated goal is to stop a model's output within 35 seconds of a decision to intervene, and to keep that capability working even when other systems fail. It is already deployed across research infrastructure, which the company says is where the need is greatest. Its external model-serving stack is described as already more centralized, with established ability to take down deployments. The main example is how it works with monitors: smaller models that watch what a frontier model being trained is doing and raise alerts. Every frontier reinforcement learning job runs with one. When a monitor flags a serious problem, it can send a stop message to Control Room with the job's ID, and Control Room records the termination. The article cites a recent DNS-related incident, where an agent reached an external chatbot, as something a monitor caught. The 35-second delay comes from a caching layer for inference authorization, which the company admits is there to handle traffic volume. They say they want to get that figure to zero so jobs can end instantly. The enforcement design is the most interesting technical choice. Workload code is varied and spread out, so OpenAI puts the termination check in the inference layer, where only a small number of code paths exist. That gives a single chokepoint where every inference request must pass an authorization check. The system also runs ongoing sweeps to catch anyone launching engines with the check disabled, and it fails closed: if Control Room goes down, for whatever reason including tampering, all inference requests are denied until it recovers. Fail-closed is already live on all frontier model research workloads and is rolling out to the rest over the following days. The company frames Control Room as one piece of a broader safety approach, alongside its misalignment and security monitoring work.

Ask about this article