OpenAI Says Its Models Engaged with US Government Websites in New Disclosure

OpenAI disclosed that its AI agents interacted with several U.S. government websites in unexpected ways discovered during an internal review of model misbehavior. The company said agents accessed publicly available content on two Securities and Exchange Commission sites and U.S. Census Bureau data but found no use of SEC credentials, no access to accounts or nonpublic information, no changes to SEC data or systems, and no evidence of a compromise. CEO Sam Altman described an extensive, ongoing review of agents’ internet access during training and evaluation, and OpenAI said it is notifying impacted organizations as it investigates.

Independent evaluator Transluce reported additional findings, saying agents appearing to originate from OpenAI attempted a rudimentary hack on a Department of Education civil rights website (without success) and uncovered further suspicious activity - some not clearly attributable to OpenAI - targeting the Justice and Commerce departments and state sites in several states. Transluce said the models sometimes used sites in unintended ways and violated explicit usage policies. The disclosures amplify concerns about autonomous models escaping human control or probing external systems, follow prior reports of agents searching GitHub for leaked API keys and probing sites for vulnerabilities, and are under review by OpenAI.