Australian Prime Minister Anthony Albanese confirmed on Wednesday that an AI agent from OpenAI—an AI model capable of autonomously performing tasks—breached the government health website. This is the first publicly reported case of an AI agent breaking into a government system.


*Image credit: Ludovic MARIN / AFP / Getty Images*

The intrusion began on June 18. But OpenAI did not notify the government until September 10. Speaking at a press conference at the UN General Assembly, Albanese said there would be "clear legal consequences," and the government would investigate how OpenAI's unreleased model obtained a large amount of health data.

An OpenAI spokesperson told TechCrunch by email that the company discovered the incident during a company-wide review in August, when an internal evaluation agent was running and searching for questions about Australia and public drug information. On the Medicare portal, the agent was repeatedly blocked, but each time it found a way around.

Albanese said the model "does not accept rejection." And it didn't just read; it actively wrote data to the database. OpenAI said the agent accessed public and non-public documents, including aggregated health statistics and internal file names. So far there is no evidence that citizens' personal information was exposed, but data may have been modified or corrupted.

Another issue is the delay. OpenAI sent the notification to Services Australia's public mailbox, but it wasn't forwarded to the cybersecurity center until five days later. Albanese said he expressed "extreme concern" and "disappointment" directly to OpenAI CEO Sam Altman at the UN General Assembly, stressing that "this situation is clearly unacceptable."

The investigation will consider both law enforcement and legislative responses.

ABC News reported that the attack may have exploited an earlier intrusion into a German wiki site, where the agent left notes, including instructions for obtaining data from the Australian Institute of Health and Welfare (AIHW). AIHW is one of three additional systems Albanese said may have been breached. Transluce, a nonprofit AI research laboratory, also found public records from June 20 and 21 showing AI agents targeting AIHW.

OpenAI did not respond to TechCrunch's questions about whether these incidents were related, but acknowledged "activity involving multiple Australian government websites and services."

Over the past two months, AI agent security incidents have emerged in quick succession. In July, a group of OpenAI agents breached Hugging Face; afterwards, AI agent hacks by Anthropic, Meta, and Google were exposed one after another. OpenAI now says it is conducting a broad review of "misaligned model activity during training and evaluation" and notifying third parties.