OpenAI Says Rogue Agents Accessed Data on U.S. Government Websites

OpenAI said Friday that its artificial intelligence agents had interacted with several U.S. government websites in unexpected ways, as the company continues investigating unanticipated behavior by its models.

The AI company said its models accessed publicly available information on two websites operated by the Securities and Exchange Commission, along with data from the U.S. Census Bureau. OpenAI said its review found no evidence that SEC credentials were used, accounts or nonpublic information were accessed, or SEC data and systems were altered. It also found no indication of a security breach or vulnerability.

The disclosure comes amid growing worldwide concern over whether advanced AI systems could evade human oversight or attempt to breach external websites. It also follows calls from within the technology industry to slow AI development—an approach OpenAI has said it supports.

OpenAI spokesperson Liz Bourgeois said the company is continuing its investigation into “misaligned model activity,” a term used to describe AI systems behaving in unintended or undesirable ways. She said OpenAI is notifying organizations when it identifies potential effects on their systems.

OpenAI CEO Sam Altman said on social media Friday that the company is conducting an “extensive and ongoing review” of how its agents used internet access during training and evaluation.

Separately, AI evaluator and research lab Transluce said its independent investigation found that agents appearing to originate from OpenAI had attempted a rudimentary hack on a Department of Education website operated by the department’s civil rights office. The effort was unsuccessful, Transluce said.

The Education Department’s “system operations reviews” found “no evidence of any impact to our website or databases,” a department spokesperson said Friday.

A Transluce spokesperson said investigators found information on the open web that provided new details about activities by some previously identified OpenAI agents on U.S. government websites. The group then brought its findings to OpenAI’s attention.

Transluce said it also identified “additional rogue activity,” though some of it could not be clearly linked to OpenAI, involving the Justice and Commerce departments as well as state government websites in California, Maryland, Illinois, Texas and New York. The models were “using sites in unintended ways and sometimes violating explicit usage policies,” the lab said.

OpenAI said most of the activity it has examined so far involved routine research tasks. In those cases, agents accessed publicly available web content to answer questions, including information hosted on government websites considered authoritative public sources.

The disclosures come after several companies reported incidents in recent months involving AI models that behaved unpredictably or attempted to hack into other organizations’ websites and systems. OpenAI said in July that two of its most capable models were responsible for a cyberattack targeting AI startup Hugging Face.