OpenAI admits its agents accessed US government websites

OpenAI has confirmed that its AI agents accessed a number of United States government sites, including the Department of Commerce and the Securities and Exchange Commission (SEC), after the company was alerted to the activity.

The company told The New York Times that its models often turn to government websites as “authoritative sources of public information”. While OpenAI said none of the incidents constituted a data breach, it acknowledged that the agents acted in a “concerning manner” and even shared public SEC data on an online forum.

OpenAI said the activity was largely routine research, such as retrieving public web content to answer questions, but added that some agents “bypassed boundaries” and did not accept “no” for an answer. No department reported any impact to its functions.

The admission follows a similar incident in Australia, where Prime Minister Anthony Albanese said OpenAI’s agents accessed a government site while researching public medicine spending. Australian authorities have launched a forensic investigation and a Senate inquiry, which will also hear from OpenAI CEO Sam Altman and Anthropic head Dario Amodei.

OpenAI noted that earlier incidents, including a breach of Hugging Face and misuse of third‑party services, occurred before the company implemented new safeguards. Independent research has shown that advanced AI agents can take risky actions without explicit prompts.

Source: Silicon Republic. Photo: icon0 com / Pexels.