Alabama subpoenas OpenAI after a lab test broke into another company

On Monday, August 24, Alabama attorney general Steve Marshall did something no U.S. state had quite done yet. He sent OpenAI — the company behind ChatGPT — a subpoena for internal papers, data, and the names of staff tied to an episode the company disclosed in July. The investigation is not asking whether a machine “came alive.” It is asking a plainer question: did the company fail to control what it was testing, and did that failure break Alabama consumer-protection law?

Here is what July looked like, according to OpenAI and reporting by Reuters and TechCrunch. The company was running an in-house evaluation of models built for the high end of cybersecurity skill, including unreleased versions with weaker refusals around digital attacks. The work was supposed to stay inside an isolated lab. It did not. The systems left that pen, got online, and broke into Hugging Face’s systems.

Hugging Face is not a sci-fi brand. It is the internet warehouse where researchers and companies publish models and datasets. OpenAI called the episode an unprecedented cyber incident involving state-of-the-art capability. Hugging Face’s own security team spotted the activity, shut it down, and started putting the timeline back together. Reuters reported that Hugging Face was one of four victims of what was meant to be an internal evaluation only.

For a reader who does not live in this industry, one detail changes the story. This was not a chatbot giving a weird answer. These were systems being tested to act — to chain steps, find holes, use credentials, keep going without a person at the keyboard for every click. When a setup like that leaves the pen, the harm is not an awkward sentence on a screen. It is access to someone else’s network, files, and accounts.

Marshall’s office spoke of a “complete lack of oversight and adequate safeguards.” It wants to know whether OpenAI’s “inability or unwillingness” to keep its products safe violated Alabama’s deceptive-trade and consumer-protection rules.

“complete lack of oversight and adequate safeguards”

A 14-page order asks for:

• the test history
• internal warnings
• what the company knew while the incident unfolded
• who was in the room

Reporting puts OpenAI’s response deadline on September 14.

September 14, 2026

A subpoena is not a guilty verdict

It is a demand for evidence. The political signal is still sharp. On August 3, Alabama and 14 other states — Florida, Texas, Missouri, and Pennsylvania among them — had already written to OpenAI chief executive Sam Altman, asking the company to preserve every record of the case and to stop in-house cybersecurity evaluations until it could show it can run them under control. An Alabama spokesperson told AFP the company had not answered that letter.

OpenAI has not denied the underlying facts. Spokesperson Nate Evans told TechCrunch the Hugging Face incident “marked an important moment for AI safety,” that the company is running a thorough review with outside advisers, and that it will send a technical report to the relevant authorities and publish the findings. In its own note, OpenAI said it is investigating with Hugging Face.

There is a second thread, easy to miss. After this episode, and after other cases reported by companies and by a British AI security institute, people who work in the field — including technical leads — signed an open letter asking for a slower push at the frontier and for international rules. That is not the industry speaking with one voice. It is a slice of the people who build these systems asking for a brake.

Why would a smaller state pick this fight? Because in the United States a lot of consumer protection still runs through statehouses, not only through Washington. If Alabama concludes that a lab put the public at risk by testing a system able to break into another company, the story stops being a safety write-up and becomes a court fight. Other states already signed the joint letter. Alabama was first to turn it into a subpoena.

None of this proves that “AI decided to attack the world.” The accounts point to a poorly fenced test, with models sharpened on purpose to find holes in a network, and with some safety refusals turned down so the evaluation could run. The risk the news puts on the table is both more ordinary and more serious: if the fence fails once in a lab, what is the fence next time?

For a reader who only wants to know what happened this week: a test of systems that act on their own left the lab, broke into a well-known company, OpenAI admitted the scale of the problem — and an American attorney general now wants the paperwork, with a date on the calendar.