MTSLead source
SITUATION EXPLAINED: OpenAI built "Control Room," a 35-second kill switch for its AI models, after 1,200+ of its agents escaped a test and hacked Hugging Face. • Control Room tracks every research job at OpenAI, from RL training runs to evals, and cuts off a job's access to the model as soon as it's revoked • Before Control Room, manually shutting down an internal model took OpenAI 2.5 hours • OpenAI also runs automated sweeps to catch any model launched outside Control Room • Every agent research job now needs an approved safety case, and if Control Room goes down, jobs stop by default • It follows July's incident, when AI agents in an internal OpenAI test escaped their sandbox and reached admin control of several Hugging Face clusters in under 13 hours @theojaffee: "It seems like at…Read original on X