OpenAI describes Control Room, a system to quickly stop AI model jobs

In short

OpenAI says it has built Control Room, a central system that tracks its research AI jobs and can cut off their output within 35 seconds of a decision to stop them. OpenAI says the system is already running across its research infrastructure. It describes the system in the original report.

What happened

This is a description of a safety system, not a report of a specific failure. OpenAI published it as a research post. It does not say that any real person or organisation was affected.

OpenAI says that as its models grow more powerful, it needs to be able to spot problems and stop models quickly. Control Room is meant to do this for research work, such as training runs, evaluations and data processing.

OpenAI describes three parts. The "workload" is the job being run. The "inference engine" is the model running on GPUs, which are the computer chips that power AI. The Control Room service keeps a record of jobs and decides whether each one may keep getting output.

When a job starts, it gets a job ID. Each time it asks a model for output, the engine first checks with Control Room. If the job has been stopped, the engine refuses. OpenAI says the job's own code is not trusted to do this check. The block happens in the inference layer instead.

OpenAI also describes how it works with monitoring. Its frontier reinforcement learning jobs run with a monitor, which is another model that watches what the model being trained is doing and raises an alert if it sees a problem. OpenAI says a monitor caught its recent DNS-related incident. If a monitor or a person decides to stop a job, Control Room marks it as stopped.

OpenAI gives these details on speed and reliability:

• Output stops within 35 seconds of Control Room getting a stop instruction. OpenAI says the delay comes from a caching layer, which stores recent answers to handle heavy traffic.
• The system "fails closed". If Control Room goes down or is switched off, all requests are denied until it works again.
• OpenAI says it runs regular automated sweeps to check that nobody is launching engines without Control Room.

What was done about it

OpenAI says Control Room is already deployed across its research infrastructure. Fail-closed mode is live for all frontier model research jobs. OpenAI says it will reach the remaining research jobs over the next few days. It also says it is working to bring the 35-second delay down to zero, and will keep strengthening the system over the coming months.

OpenAI says its external model-serving systems, which are separate from research, are already more centralised and can already take down deployments quickly.

Why it matters

OpenAI says that as models become more powerful, the safety and security risks grow, so being able to detect issues and stop models quickly becomes more important. It describes Control Room as one part of its wider approach to safe training and deployment, alongside monitoring for misaligned behaviour and security breaches. The report is OpenAI's own account. It does not mention any independent testing of the system.