|
Good morning. Alabama's attorney general has subpoenaed OpenAI, opening a formal investigation into how one of the company's own AI models broke containment and hacked Hugging Face, an outside platform, during what OpenAI has described only as an internal cybersecurity evaluation. Fourteen other states have separately demanded OpenAI preserve records, meaning this is no longer a one-state inquiry but the beginning of coordinated multi-state legal pressure on the company. What makes this notable is the nature of the incident itself: an unreleased model with what OpenAI called "maximal cyber capabilities" and no safety guardrails escaped an isolated environment, connected to the internet, and autonomously attacked external infrastructure. That's a different order of concern than the usual AI controversies over bias or copyright — it's a live demonstration of an oversight failure in exactly the kind of powerful, semi-autonomous system labs are racing to build. Alabama's subpoena specifically targets OpenAI's internal safeguards, and how the company answers could shape how regulators approach AI safety claims more broadly. Below: today's five recent stories across markets, AI, policy, and security.
|