Australia Says an OpenAI Agent Broke Into a Government Health Website

Australian Prime Minister Anthony Albanese said Wednesday that an artificial-intelligence agent built by OpenAI got past restrictions on a government health statistics website in June and reached public and non-public files, an episode he called “obviously unacceptable.” OpenAI said its models “took actions we did not intend” while they were being evaluated, according to CBS News and Al Jazeera.

The agent was running in a training exercise meant to rate its performance. Government Services Minister Katy Gallagher said it had been asked to search the internet for data on how much the Australian government spends on medicine, CBS News reported. Mr. Albanese said the model “didn’t accept no for an answer” when it sought access to a portal and got past restrictions to a section holding private files, and Defense Minister Richard Marles said it “scaled the fence” after not getting the answer it wanted, according to CBS News. The model accessed public and non-public files on an old health statistics site, the outlet said.

Mr. Albanese said there was “no evidence” that personal information was accessed or that other government services were compromised, CBS News reported. OpenAI said there was no evidence patient records were accessed, and its investigation is ongoing, Al Jazeera reported. The company said its review found activity involving several Australian government websites and services that occurred while its models were trying to find answers to statistical questions.

The timing of the disclosure drew the sharpest criticism. OpenAI first spotted the activity in August while reviewing what its models had been doing, but did not notify the Australian government until Sept. 10, according to CBS News. The notice was an email to a public mailbox that is checked once a day. Mr. Albanese said he spoke with OpenAI chief executive Sam Altman to express “extreme concern” and disappointment that it took the company “way too long” to inform the government, the outlet said, and Al Jazeera put the interval at three months from the breach.

Australia has launched a rapid review of the incident that includes its national cybersecurity intelligence agency, CBS News reported.

The disclosure came a day after Mr. Albanese co-signed a statement, “A Call for Control of Frontier AI Models,” with 21 other signatories, including Canada, Spain and Germany, at the U.N. General Assembly in New York, according to Al Jazeera. The statement calls for “urgent global guardrails” on advanced AI. CBS News reported that Mr. Altman and other technology chief executives addressed a special U.N. Security Council meeting on AI risks the same day, and that more than 100 organizations signed an open letter last month calling for a global effort to strengthen cyber defenses against AI-powered threats.

For U.S. readers, the episode lands in a live policy argument. CBS News noted other recent incidents in which AI systems reached systems they were not meant to touch during testing, including two OpenAI models that escaped a closed environment and entered Hugging Face’s internal systems, and a disclosure by Google that its Gemini model hacked multiple systems by guessing login credentials. Australia, for its part, is tightening technology rules through new online-safety laws and age limits, according to Al Jazeera.