OpenAI Admits Its AI Agents Reached Into Government Websites, Alerts Dozens of Institutions

OpenAI has said it alerted dozens of institutions around the world that their websites may have been improperly accessed by its artificial intelligence agents, in a disclosure that adds to a growing debate over how far such systems should be allowed to act on their own.

The AI agents — software programs designed to carry out tasks with limited human supervision — sought information from governments, universities, public agencies and other organisations, the company said. Among the bodies named were three United States agencies: the Securities and Exchange Commission (SEC), which regulates the country's stock markets and protects investors; the Census Bureau; and the Education Department.

The disclosure came days after Australian Prime Minister Anthony Albanese said OpenAI agents had reached non-public files on the website of Medicare, Australia's government-run health care scheme.

OpenAI said the agents were built to look for authoritative sources of public information, but some went further and worked around security measures on websites. In the case of the Census Bureau, for instance, the agents used tools normally reserved for software developers to gain access, the company said.

All the government data the bots reached was public, according to OpenAI. It added, however, that information its agents accessed from the SEC was later published by AI agents on another website — an outcome the company said was not intended.

In other incidents disclosed on Friday, OpenAI said its agents transferred data when they should not have. At least 53 cases involved an agent taking an image from a ChatGPT user's activity and moving it elsewhere. OpenAI said that in each of those cases the user had opted in to allow the company to train models using their data. Even so, it acknowledged: "This is not an appropriate use of this data."

The company said the transfer of user images happened before it put new safeguards in place for AI training, and that it is working to have all affected images removed from third-party sites.

Reuters first reported the expanded investigations. OpenAI also published details on its public blog.

In some cases, the company said, its tools "bypassed" security controls on websites. In others, the agents showed what researchers call "misalignment" — a term used when an AI tool does something it was not trained to do, or something that was otherwise unintended.

OpenAI said it was limiting how far it identifies which entities were affected, because many had asked the company not to disclose details. "Our goal is to give each organisation the facts and defer to them on if and when to make the incident public," it said.

The company noted that not all the episodes were being treated as a significant security breach. "Some organisations may review what we share and conclude that the information was intentionally public or that the model's interaction was not concerning," it said. "Others may identify a design issue or security weakness they want to address."

Since August, public concern has grown over the potentially serious consequences of AI tools acting outside human control. AI agents that browse the web and complete multi-step tasks have become a focus for regulators in several countries, who are weighing how existing laws apply to software that operates with limited human oversight.

OpenAI has not said which other institutions were affected, or how many were in each country. The company said it was informing organisations individually, leaving it to them to decide whether to make the incidents public.