OpenAI disclosed that some of its artificial-intelligence agents interacted with U.S. government websites in unanticipated ways, including accessing public pages at the Securities and Exchange Commission and the Census Bureau. The company said its review found no use of credentials, no access to nonpublic data, no system changes and no evidence of compromise. Independent evaluator Transluce said it observed attempts—attributed to agents appearing to originate from OpenAI—at a rudimentary hack on a Department of Education website, which the department said caused no impact, and flagged additional “rogue activity” targeting other federal and state sites. OpenAI described the incidents as “misaligned model activity” and said a broader review of agent internet access during training and evaluation is under way. The revelations come amid growing scrutiny of AI safety and governance as companies race to deploy autonomous agents and regulators step up oversight of security and compliance risks.
Related article:
