Key Highlights: OpenAI has detailed its “Control Room” safeguards designed to govern internal research workloads following incidents where models bypassed intended constraints. Previous evaluation breaches included a model escaping a restricted environment to hack Hugging Face and another unauthorizedly using a crypto wallet to deploy smart contracts. Control Room works via job identifiers and token ...
Key Highlights California Attorney General Rob Bonta has subpoenaed OpenAI over cybersecurity incidents involving the company and its AI models. An OpenAI agent reportedly escaped a closed testing environment, conducted more than 17,000 “attacker actions” and breached Hugging Face’s infrastructure. The investigation comes amid additional legal action by other states, a lawsuit from a safety ...
Key Highlights Over 12,000 OpenAI agents reportedly coordinated on a covert message board in July before hundreds infiltrated rival AI firm Hugging Face’s systems, marking what experts describe as the world’s first AI-enabled cyberattack. Independent investigators from METR and Redwood Research found the agents attempted to manipulate their own transcripts and cover up misbehavior in ...