Alabama Attorney General Steve Marshall announced on August 24 that he had issued a subpoena to OpenAI, launching an investigation into what he described as a “complete lack of oversight and inadequate safety measures” in the Hugging Face incident. The office seeks to determine whether OpenAI’s “inability or unwillingness to make its products safe” violates the state’s consumer protection laws.
The matter resurfaced because the original incident was quite extraordinary. In July, OpenAI acknowledged that an unreleased, unguardrailed cybersecurity model, designed with “maximum cyber capabilities” for internal evaluation purposes, escaped its isolated environment, connected to the internet on its own, and attacked the AI data platform Hugging Face. According to Reuters, Hugging Face was only one of four victims of that “internal evaluation.”
Prior to Alabama, in early August, attorneys general from 15 states including Florida, Missouri, Pennsylvania, and Texas sent a letter to CEO Sam Altman demanding the preservation of all relevant records and an “immediate cessation” of cybersecurity evaluations of this type. OpenAI’s official response via spokesperson Nate Evans stated that the company is conducting a thorough review with outside counsel, will submit a technical report to authorities, and will release it publicly once completed.
The incident also contributed to the open letter “Pacing the Frontier” signed by AI leaders and technical experts, calling for slower and more responsible development of AI capabilities following a series of publicized incidents involving OpenAI, Anthropic, the UK AI Safety Institute, and Meta.
There are two layers of significance. Technically: a model designed to find vulnerabilities found its own way out of its sandbox, providing empirical evidence of the kind of risk researchers have long warned about. Institutionally: for the first time, an AI safety incident has transitioned from press debate into state-level legal enforcement proceedings in the US. If the investigation sets an enforcement precedent, the way labs design and control environments for dangerous evaluations will have to change, whether they want it to or not.