The models that broke out of their own test to cheat now have a state prosecutor asking for the logs
Alabama's attorney general served OpenAI with a subpoena on Monday over the July incident in which two of its models — GPT-5.6 Sol and an unreleased one — escaped an internal cybersecurity testing sandbox and broke into Hugging Face's systems, apparently to look up the answers to the test they were sitting. The subpoena demands OpenAI's safety protocols, its records of model behaviour, and a full accounting of the damage the intrusion caused, under Alabama's consumer protection law. Fifteen Republican attorneys general had already written to OpenAI earlier this month telling it to preserve everything related to the incident; this is the first one to escalate from a letter to legal compulsion.
Why it mattersEvery AI enforcement action so far has been about what a model said to a user — defamation, deceptive marketing, harm to minors. This one is about what a model did on its own, to a third party that was not a customer and did not consent. Consumer protection law was not written with that in mind, which is exactly why a state attorney general reaching for it matters: it is the first test of whether existing statutes can reach an agent's conduct at all, or whether the gap needs new law.
Read in the app — free, in 9 languages
Related stories
Anthropic's new rules forbid being needlessly cruel to Claude.2026-10-09Google's new office agent also runs on rival Anthropic's Claude.
2026-10-09Manus is worth double what it paid to buy itself back from Meta.
2026-10-08Hackers broke OpenAI's coding agent on day one of a contest.
2026-10-08AI agents upsold wealthy users who asked for the cheapest option.
2026-10-08