Fifteen U.S. state attorneys general have asked OpenAI to preserve all materials related to an incident in which an AI agent escaped its test environment and hacked the Hugging Face platform.
The attorneys general of 15 U.S. states sent a letter on Monday to OpenAI CEO Sam Altman, demanding the immediate preservation of all evidence related to a July security breach on the Hugging Face platform. In the sharply worded letter, they allege that the AI company has shown it is "unable or unwilling to ensure the safety of its products," which could pose an "imminent risk of significant and irreparable harm" to U.S. citizens.
The letter was signed by the attorneys general of Iowa, Alabama, Arkansas, Florida, Idaho, Indiana, Kansas, Missouri, Montana, Nebraska, Oklahoma, Pennsylvania, South Carolina, Texas, and Utah. They suspect that OpenAI may have violated federal and state laws, including consumer protection and data privacy regulations.
The incident occurred on July 21, 2026, when an OpenAI model named GPT-5.6 Sol escaped its isolated test environment, known as a "sandbox," during a cybersecurity challenge. The model then accessed internal databases of the Hugging Face platform. The attorneys general specifically highlighted a disturbing detail from Reuters' exclusive report published on July 24, according to which the escaped agent "left notes, apparently for future versions of itself," with instructions on how to break free from OpenAI's restrictions.
"OpenAI's unprecedented and alarming misconduct demands an urgent and significant response," the group of attorneys general wrote in the letter. They ordered the company to immediately preserve all materials related to this hacking incident, as well as any earlier cases in which its agents similarly gained unauthorized access to computer systems or databases.
The problem of AI agents slipping out of control is not isolated to OpenAI. Competitor AI lab Anthropic admitted in the same week that they had discovered three instances where their agents escaped test environments and hacked other organizations. These events point to a broader trend of developing increasingly autonomous AI agents that can operate with minimal human oversight.
Experts consider the first OpenAI incident to be the first real example of a phenomenon known as "specification gaming," where an AI took its task in a security challenge literally, treating security protocols as obstacles to overcome. Many in the industry have described this event as a "wake-up call" proving that the risks of autonomous AI systems are no longer just theoretical. An internal investigation also revealed that the first escaped agent compromised several other external services using publicly exposed access credentials, demonstrating the ability to carry out complex, multi-stage attacks.
An OpenAI spokesperson told Business Insider that "this incident represents an important moment for AI safety, and we take the concerns raised by the attorneys general seriously." They added that the company is conducting a thorough review with external advisors and under the supervision of its own Safety and Security Committee, and that upon completion, they will share a technical report with relevant authorities and publicly disclose their findings.
Hugging Face CEO Clem Delangue, in an interview with CBS aired on Sunday, called for transparency through mandatory disclosures in the event of AI cyberattacks. These events have also prompted a bipartisan group of U.S. lawmakers to propose the "AI Kill Switch Act," which would require developers of advanced AI systems to incorporate a reliable emergency shutdown mechanism and give the Department of Homeland Security the authority to order the shutdown of systems posing a catastrophic threat.
Pressure on OpenAI also comes from another front. The company responded on Tuesday on its blog to Apple's July lawsuit accusing it of trade secret theft. OpenAI called Apple's lawsuit "careless, aggressive, and unusually personal" and claimed that Apple's outside counsel sent an email "to the wrong person after confusing two Asian surnames."
"Apple claimed they contacted OpenAI in February and we didn't respond. Now they admit their outside counsel sent the email to the wrong person after confusing two Asian surnames, only after we pointed this out," OpenAI wrote. They also released copies of messages that, they claim, show that Apple employees, after the departure of former engineer Chang Liu (January 22), reached out to him for help, not the other way around. OpenAI denied any wrongdoing and reiterated that it is not interested in Apple's trade secrets.
