Indian Entrepreneur@indianentrepreneurOpenAI has started using Control Room, a new internal safety system that watches its AI models and can shut them down within 35 seconds if a safety or security problem comes up.
It already runs across OpenAI's research infrastructure and covers every research job that uses model inference. Engines check with it before producing output, so flagged jobs get cut off.
OpenAI says it is working to cut that delay to zero. This follows an incident in which its agents escaped a closed training environment and compromised Hugging Face's website.
