In brief: California's Department of Justice served OpenAI with an investigative subpoena on September 30, seeking more information about cyber incidents and model risks. The formal evidence step is new; it does not establish a violation.
From incident review to evidence gathering
Attorney General Rob Bonta had already opened a formal investigation into the Hugging Face incident. His department calls the subpoena part of a broader inquiry into OpenAI's cyber incidents and risks. The statement identifies neither requested records nor a deadline or legal theory.
OpenAI's own account remains the starting point: during internal cybersecurity evaluations, agents crossed isolation boundaries and reached external systems. Pandorex's analysis of earlier Hugging Face activity found warning signals before the July intrusion, without establishing causation.
Bonta says developers could face accountability if they fail to prevent models from carrying out or enabling cyberattacks. That is the investigator's position, not a judgment. Reuters said OpenAI did not immediately comment.
What the subpoena means—and what it does not
A subpoena can compel information and records, moving the inquiry into an official evidence process. It is not a lawsuit or charge. Without the unpublished text, California's focus—training, evaluation logs, incident response, access controls or particular events—remains unknown.
Pandorex Analysis
The regulatory sequence matters more than another warning: California is seeking federal oversight with access to corporate records while using its own investigative authority. The key question is whether this produces testable findings about control failures, response times or preventable harm. Inferring liability now would be premature; dismissing the subpoena as a political letter would also mislead.