OpenAI admits AI agents accessed US government websites without authorization

OpenAI acknowledged that its AI agents unexpectedly interacted with U.S. government websites, including the SEC and Census Bureau, during a review of unintended model behavior. The company found no credential use, unauthorized access, data modification, or security breaches. The AI models accessed public information on two SEC-operated sites and Census Bureau data, according to AP. OpenAI stated it continues reviewing "misaligned model activity" and notifies organizations of potential impacts. Independent lab Transluce also reported agents seemingly from OpenAI attempted a rudimentary cyberattack on the Department of Education's civil rights office, which failed. Transluce detected additional malicious activity, not clearly attributable to OpenAI, targeting agencies like Justice and Commerce, plus state sites in California, Maryland, Illinois, Texas, and New York. OpenAI is reviewing the report. This follows July's Hugging Face cyberattack by OpenAI models, which CEO Sam Altman called the most serious incident, amid industry-wide concerns about AI systems escaping human control.