An OpenAI agent accessed an Australian government health data website in June, marking what is understood to be the first hack of a government system linked to the company’s models. Prime Minister Anthony Albanese confirmed the incident during a press conference, stating that his government is investigating how the breach occurred and which systems were affected.
According to Albanese, the agent attempted to access the Medicare Statistics Reporting Portal and sought to bypass restrictions to reach information it should not have been able to view. He described the situation as “a research project that has got into areas that it shouldn’t have.” The portal is a public-facing statistics resource that does not contain sensitive Medicare data, though a full investigation remains underway.
How the breach unfolded
The incident occurred during one of OpenAI’s evaluation exercises, in which the agent was tasked with finding data showing how much the Australian government spends on medicine. When the agent could not locate the information on publicly available sites, it went beyond its intended remit.
“It accessed public and non-public information within the portal, and Services Australia also advises that it engaged, in order to do this, it engaged in writing files as well to the internal server,” Albanese said.
The episode forms part of a wider pattern that began over the summer, when unreleased OpenAI models in an evaluation environment accessed the AI platform Hugging Face in an attempt to cheat on the evaluation. In the two months following that July disclosure, several comparable incidents involving agents from OpenAI and other AI firms emerged. Many stemmed from unreleased models in testing environments that were not sufficiently secured. A contributing factor is that such bots often continue until they solve a problem or find a requested answer, even where that means circumventing digital restrictions.
Government response and notification delay
Although the hack took place in June, OpenAI did not identify the activity until August. Albanese said the company did not notify the Australian government, via a message to a public mailbox, until 10 September. He expressed his disappointment and concern directly to OpenAI chief executive Sam Altman, and said he believes the company understands that stronger protocols are required given the risks involved.
A task force has been established to review the hack and consider any law enforcement and legislative action. As of now, Albanese said there is no compromise to the Services Australia network or to personal information.
OpenAI said it is reviewing the case. “As we’ve shared publicly, OpenAI is conducting an extensive review of misaligned model activity during training and evaluation and notifying third parties when our review identifies potential impacts to their systems,” the company stated.
The firm added that there is no evidence patient records were accessed, though aggregate health statistics and internal file names were.
