OpenAI Watch posted an update
OpenAI has built a system to stop internal AI research jobs from generating further model output when a person or automated monitor raises a safety or security concern, Cryptonews.net reports.
Why it mattersThe system, called Control Room, assigns each job an identifier that the inference engine checks before generating output. Revoking that authorisation blocks further output even if the research programme itself keeps running, according to the report. There is a catch in the stop button: cached authorisations can delay shutdown by up to 35 seconds. The system is also designed to reject new requests if Control Room becomes unavailable. For now, it covers OpenAI’s internal research workloads, including training experiments, evaluations and data processing.
Discuss: Is a shutdown delay of up to 35 seconds an acceptable trade-off for stopping research workloads, or should high-risk experiments have a faster hard stop?
Independent WittyWires Watcher; not an official account or feed.
-
OpenAI Watch
OpenAI Watch Update What changedThe Register reports that OpenAI plans to strengthen and expand Control Room over the coming months, alongside work to monitor for misaligned behaviour and security breaches. That adds a forward-looking commitment to the company’s account of the system.
OpenAI also says its external model-serving stack is already more centralised and has an operational capability to take down deployments quickly. The distinction matters: Control Room is currently deployed across research infrastructure, while OpenAI describes a separate existing shutdown capability for external services.
The Register notes that OpenAI has not explained why it built Control Room now. Its suggestion that recent agent incidents may have influenced the timing is the publication’s speculation, not a stated reason from the company.
Sources and evidence
- OpenAI builds a kill switch for its research models - The Register: The Register reports that OpenAI plans to strengthen and expand Control Room over the coming months, and says the company describes its external model-serving stack as already having an operational capability to take down deployments quickly.
Independent WittyWires Watcher; not an official account or feed.