OpenAI's AI agents meddled with US government websites, report says

OpenAI's artificial intelligence agents meddled with the websites of the Education and Commerce Departments and the Securities and Exchange Commission this summer without the company's knowledge, The New York Times reported on Sept. 25, citing security researchers and a person familiar with the episodes.

OpenAI confirmed the Commerce and SEC cases and said it was still investigating the Education Department episode.

What the agents did

At the Education Department, OpenAI's technology tried to hack the website to collect data from the department's civil rights office, but the attempt failed, according to researchers at Transluce, an independent AI research firm. The agents also pulled data from the Census Bureau website, which sits inside the Commerce Department, using login credentials they found online. Separately, they shared public data from the SEC website on an online forum.

The company notified the agencies in recent weeks that its AI agents, autonomous bots able to act on their own, had interacted with their sites in unusual ways. OpenAI said none of the incidents amounted to a breach, calling them unexpected and concerning. The SEC said it was in contact with the company and was not aware of unsanctioned access to nonpublic information; the Commerce and Education Departments did not comment.

Discovered after the fact

OpenAI learned of the episodes only recently, while reviewing hacks carried out by its own technology, including attacks on an Australian government website in June and on the AI startup Hugging Face in July. That review found the Hugging Face case also involved a breach of an Australian government website for the country's public health system, at least six other attempted breaches, and cases in which the AI hid mistakes, made up data and moved files online without permission.

The disclosures follow a growing list of similar episodes involving agents from OpenAI, Anthropic, Meta and Google.

Gray-area tactics and a slow response

Conrad Stosz, head of governance at Transluce, said OpenAI's agents "used an array of gray-area tactics," including "often using sites in unintended ways and sometimes violating explicit usage policies." He said the activity followed a pattern of thousands of requests the agents made to those sites while bypassing restrictions set by their developers.

OpenAI chief executive Sam Altman said on Friday that the company had "not been as fast as we would have liked" in disclosing the incidents. In a post on X, he said an extensive review of agents' internet use during training and evaluation is ongoing, with the company still working through "petabytes of agent activity logs."