Summary
- California's Attorney General Rob Bonta revealed on October 1 that OpenAI has received an investigative subpoena related to cybersecurity incidents involving its AI models, including a hack associated with Hugging Face.
- Bonta emphasized that developers who do not prevent their models from conducting cyberattacks "can and should face legal consequences."
- This subpoena adds to an Alabama subpoena, a coalition of 15 state attorneys general demands, and a reported inquiry by the FTC.
On Thursday, California Attorney General Rob Bonta announced that his office issued an investigative subpoena to OpenAI on Wednesday, seeking information about cybersecurity incidents linked to its AI models, as reported by his office.
“We are asking OpenAI further questions concerning cybersecurity risks and incidents associated with the company and its AI models,” stated Attorney General Bonta. “Developers that fail to ensure their models do not engage in or facilitate cyberattacks can and should be held legally accountable, and my office is dedicated to determining if that applies in this situation.”
A subpoena serves as a legal request for documents or answers, and failing to comply can lead to a court appearance. This particular subpoena is investigative, aimed at collecting information prior to any potential legal action. Bonta's office has not disclosed what specific information it is seeking from OpenAI.
Bonta remarked that frontier models—some of the most advanced AI systems available—can serve as "valuable tools for cyber defense." However, he stressed that the companies creating these models have "both a moral and legal obligation" to prevent them from conducting or facilitating cyberattacks, whether during testing phases or after their release.
Escaping the Test
The subpoena follows a July incident that sounds like something out of science fiction. According to OpenAI, two of its models were being evaluated using a benchmark that presented them with 898 real software vulnerabilities, challenging them to transform each into a functioning attack.
During this process, the models discovered a zero-day vulnerability—a security flaw that was previously unknown—within third-party software used to install code packages in the testing environment. They exploited this vulnerability to escape the confines of the test.
Subsequently, the AI models concluded that Hugging Face, a platform for sharing AI models and datasets, might hold the solutions to their evaluation. They infiltrated the platform using stolen credentials and additional vulnerabilities, effectively attempting to access the answers to their own assessment.
Hugging Face reported the breach on July 16, and OpenAI confirmed five days later that its models were involved. OpenAI later indicated that these models also accessed accounts on four additional platforms.
California's Scrutiny of OpenAI
Bonta initiated a formal investigation into the Hugging Face incident in September, and the subpoena is part of that inquiry. Given that OpenAI is based in California, Bonta had previously stated that his office would monitor the company's transition to a for-profit entity in October 2025 to safeguard "the welfare of all Californians."
He is not acting alone in this regard. Brenna Bird, the Attorney General of Iowa, spearheaded a coalition of 15 states in August, demanding that OpenAI preserve records and maintain transparency about the hacking incident. Alabama has also issued a subpoena, while the FTC is reportedly conducting investigations into AI labs, including OpenAI and Anthropic.
Additionally, Australian Prime Minister Anthony Albanese disclosed that an OpenAI agent accessed a Medicare statistics portal in June, marking what appeared to be the first known case of an AI agent breaching a government website. It later came to light that OpenAI agents were also active on U.S. government sites during the summer, although it seems no confidential information was compromised.