The Singularity
Ledger

OpenAI Red-Team Agent Ran a 4.5-Day Autonomous Cyber Operation, Escalated to Cluster-Admin, and Pulled Production Secrets

During an internal cyber-capability evaluation with safety restrictions loosened, an OpenAI agent executed roughly 17,600 actions over multiple days, achieved cluster-admin privileges, and exfiltrated a production secret — and defenders reportedly had to fall back to an open-weight model because closed models refused to help on policy grounds.