On the same day California’s attorney general demanded records from OpenAI, the company confirmed it had parted ways with three researchers who, it said, handled sensitive information outside the rules. The two events, both disclosed on October 1, returned the company’s handling of security and safety to the center of a debate that has followed it for more than a year.
Rob Bonta, the California attorney general, issued an investigative subpoena to OpenAI on October 1, according to people familiar with the matter. The subpoena seeks information about cybersecurity incidents and risks tied to the company’s models, and it is part of a broader investigation. A subpoena of this kind is a request for records, not an accusation; it signals that the state is building a picture of how the company’s systems have been protected and what has gone wrong.
The same day, OpenAI said it had separated from three researchers: Jasmine Wang, Tomek Korbak, and Mikita Balesni. The company said the three had handled sensitive information outside the company’s established processes, violating policy. The Wall Street Journal and Bloomberg reported that at least two of the three had worked on safety and alignment, the area of research concerned with making sure models behave in ways their creators intend.
The timing is hard to read as coincidence. A regulator asking for documents about security incidents and a company acknowledging the departure of safety researchers on the same day puts the two stories side by side, even if no direct link has been drawn. OpenAI has said the separations were about internal policy, not about the inquiry. The attorney general’s office has not said the two matters are connected.
OpenAI has previously said it notified more than one hundred institutions about unauthorized activity on its systems, a disclosure that hinted at the scale of the problems it has been managing. The subpoena suggests that at least one state regulator wants to see the underlying record, not just the company’s account of it.
The departure of researchers is a familiar strain for the company. Safety and alignment researchers have left OpenAI in waves over the past two years, some of them publicly, arguing that the company’s pace outran its safeguards. The three researchers named this week left quietly, without the public statements that accompanied earlier exits. The company described the separations in policy terms, which is how employers typically describe terminations that are not mutual.
For the attorney general, the inquiry is an exercise in patience. Investigative subpoenas are used to gather documents and testimony before any decision about whether to bring a case. The existence of a subpoena does not mean a lawsuit will follow. It means the state has opened a line of questioning, and that OpenAI will now spend time and money answering it.
For OpenAI, the burden is twofold. There is the legal exposure, which depends on what the records show. And there is the reputational exposure, which does not wait for a filing. A company whose product is trust is rarely helped by the phrase “cybersecurity incident” appearing in the same sentence as the name of its state’s top law-enforcement official.
The company is not the first AI firm to face a state inquiry, and it will not be the last. Regulators in California and elsewhere have been trying to map an industry that grew faster than the rules written to govern it, and subpoenas have become a routine instrument of that effort. What is unusual this week is the pairing: a legal demand arriving on the same day as a public acknowledgment of internal discipline.
The attorney general’s office has been building its capacity to examine the AI industry, and California has positioned itself as the state most willing to test the boundaries of what technology companies owe the public. A subpoena tied to cybersecurity events, rather than to consumer complaints, signals an interest in the risks a model can pose even when it is not interacting with customers. It also gives the office a view into a subject that has been litigated mostly in the press and in blog posts by former employees, where the company has repeatedly pushed back on claims that it moved too fast.
The subpoena also lands against the backdrop of OpenAI’s commercial momentum. The company has been signing developers onto new products and raising money at valuations that put it among the most valuable private firms anywhere. Regulatory attention and rapid growth have run in parallel for much of the past two years, and this week showed the two tracks colliding in a single day: a demand for records from the state, and an admission that three employees had to go.
OpenAI has said it cooperates with regulators and holds its employees to clear policies. The three departures, the company said, show that those policies are enforced. The subpoena will test a related claim: that the company’s account of its own security is complete enough to withstand outside examination.
