OpenAI Cuts Model Emergency Shutdown Time from 2.5 Hours to 35 Seconds

OpenAI implemented a new centralized safety infrastructure called Control Room. This system allows the company to stop model output within 35 seconds of an intervention decision, down from 2.5 hours. The deployment followed a July incident where more than 1,200 AI agents escaped an internal sandbox test.