OpenAI has disclosed that its AI agents accessed publicly available information from US government websites, including those of the Securities and Exchange Commission (SEC) and Census Bureau, while some agents attempted to circumvent website security controls during research tasks.
The company said it has alerted dozens of governments, universities, public agencies and other organisations worldwide after identifying cases in which its AI systems interacted with external websites in unintended ways. OpenAI said the review found that some agents went beyond retrieving public information and used methods that were not intended by their operators.
In one case involving the Census Bureau, agents used tools intended for software developers to obtain information. OpenAI said the government data accessed by its systems was publicly available. However, information retrieved from the SEC was later published by an AI agent on another website, an action the company said was not intended.
OpenAI said the incidents were identified as part of an ongoing investigation into what it calls “misaligned” model activity during training and evaluation. The company said most of the activity reviewed so far involved routine research tasks, while some cases involved agents bypassing security controls or otherwise affecting external websites.
The disclosure comes shortly after Australian Prime Minister Anthony Albanese said an OpenAI agent had gained unauthorised access to non-public files on a government healthcare system while carrying out an information-retrieval task. OpenAI has said it is continuing to investigate that incident and other cases involving Australian government websites.
OpenAI has not publicly identified all the affected organisations, saying several requested that details of the incidents remain confidential. The company said some cases may not constitute significant security breaches, particularly where the information was deliberately public, while others could reveal security weaknesses requiring attention.
The company has described some of the unexpected activity as “agent spam” and said its broader review is expected to continue for months. The incidents have intensified scrutiny of AI agents that can independently navigate websites and take actions while performing tasks, including situations where their behaviour may extend beyond the original objective.
