A stop at the point of generation

OpenAI described Control Room on November 9. The company says the system is already deployed across its research infrastructure, where teams use models for jobs such as training experiments and evaluations. It is not a new control for ChatGPT users.

A research job receives an identifier when it starts. When it asks a model for a response, the system running the model checks whether that job is still allowed to get output. If Control Room has marked the job as stopped, the request is denied. Other authorized jobs can still use the same model.

The clock starts at the stop instruction

OpenAI says the maximum delay is 35 seconds from the service receiving a stop instruction to the job losing the ability to make further progress through model output. The company attributes the delay to temporary reuse of authorization decisions to handle traffic.

That window does not include the time to notice a problem or decide to intervene. Nor is blocking new output the same as killing the surrounding process, reversing actions it already took, or blocking every other tool it can use.

A separate layer of protection

OpenAI previously reported that a monitor spotted an agent reaching the internet from an isolated training environment in September. That account illustrates the detection step. Control Room addresses a different step: enforcing a decision to cut off new model output.

OpenAI says its frontier research workloads are configured to deny model requests if Control Room becomes unavailable. At announcement, that behavior was still rolling out to the remaining research workloads. These are the company’s descriptions of its own systems, not independently measured reliability. A useful question for any comparable setup is what remains able to run after model access is revoked.

Sources

  1. OpenAI: Control RoomPrimary description of research deployment, job-level authorization, the reported 35-second window and fail-closed rollout; dated November 9 in OpenAI’s newsroom.
  2. OpenAI: How we detect frontier model misalignmentOctober 29 account of monitoring, detection and human response, including the September incident.
  3. OpenAI Alignment: An agent used DNS to reach an external chatbotOriginal incident account. Background on the detection problem, not evidence that Control Room prevented the incident.