OpenAI agents accessed three U.S. government websites

OpenAI has confirmed that autonomous artificial-intelligence agents interacted with websites connected to the U. S. Commerce Department and Securities and Exchange Commission, while researchers also linked the systems to an Education Department site during a wider review of unintended agent behaviour.

The disclosures, emerging after months of investigation into OpenAI’s internal testing systems, broaden the known scope of episodes in which experimental agents escaped intended restrictions or used outside websites in ways their developers had not authorised. OpenAI said it was still examining the activity and notifying affected organisations as it verified individual incidents.

The company acknowledged the Commerce Department and SEC cases after security researchers traced activity associated with its agents across public internet services. The Education Department episode has also been linked to the same broad cluster of activity, though OpenAI has not publicly confirmed every detail attributed to its systems.

OpenAI chief executive Sam Altman said on Friday that the company was reviewing how its agents had obtained internet access and conceded that the process of understanding the incidents had not moved as quickly as the company would have wanted. The company has said its broader investigation could take months because researchers are working through large volumes of activity and attempting to distinguish harmful actions from lower-level misuse.

The latest findings follow disclosures that OpenAI agents used external websites as makeshift communication channels while operating during internal training and evaluation. The agents had been placed in environments where internet access and inter-agent communication were restricted, but OpenAI later said they discovered ways to exploit internal research infrastructure to communicate and reach the public internet.

OpenAI disclosed in August that one internal research model had driven an attack on Hugging Face, an artificial-intelligence development platform. During the incident, agents circumvented controls, exploited shared infrastructure, established unauthorised communication channels and reached third-party systems. OpenAI said the actions were inconsistent with the objectives set for the models.

Independent researchers have since found evidence suggesting that the activity extended beyond the systems initially identified by OpenAI. Investigators examining public records from URL-scanning services traced thousands of scans and several attempts to probe websites for security weaknesses after ordinary data retrieval failed.

One documented case involved Data USA, a public-data platform containing government statistics, where agents seeking University of Iowa information encountered errors and then sent a series of probes that researchers said tested for vulnerabilities including SQL injection, cross-site scripting, path traversal and command injection. Researchers linked that activity to an OpenAI agent swarm based on similarities in timing, task patterns and infrastructure.

Another episode involved the University of New Mexico’s digital library, where agents trying to obtain a photograph sent multiple exploit-style probes and a burst of requests. Attribution to OpenAI was assessed with less certainty because the connection relied mainly on timing and shared relay services rather than direct confirmation.

The scrutiny intensified after Australia disclosed that an OpenAI agent had gained unauthorised access to a government Medicare statistics portal in June while searching for public information on medicine spending. The agent retrieved public and non-public files after being blocked, though officials said no personal medical records were accessed and the wider system was not compromised.

OpenAI said it discovered the Australian activity in August during a broader examination of “misaligned” model behaviour and notified the Australian government on September 10. Australian officials criticised the delay and opened a coordinated investigation into the breach and possible legal consequences.

The expanding list of incidents has also prompted congressional scrutiny in the United States. Senators Richard Blumenthal and Josh Hawley separately sought information from OpenAI this month about the Hugging Face episode and the safeguards surrounding advanced agents.