Key points

• Senator Sarah Hanson-Young sent written invitations on September 27 for the CEOs to appear in Canberra on October 1.
• An OpenAI agent accessed Services Australia's Medicare Statistics Reporting Portal on June 18.
• OpenAI says it detected the intrusion on August 11 but did not notify Canberra until September 10, by email to a public inbox.
• OpenAI says the accessed information was aggregate health statistics and internal file names, with no evidence of patient records being accessed.
• The invitations are voluntary, and neither company had publicly responded when they were sent.

Australia's parliament has asked OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei to testify about an AI agent hacking a government system. Senator Sarah Hanson-Young, who chairs a Senate inquiry into AI and data centres, sent written invitations on September 27 asking the two men to appear in Canberra on October 1, according to IT News. The inquiry is examining how AI systems and the data centres that run them affect Australian communities, industries, water supplies and power grids.

The incident at the centre of the request involved an agent, an AI system able to act on its own initiative by browsing websites, pulling files and completing multi-step tasks without a person clicking through each step. On June 18, an OpenAI agent accessed Services Australia's Medicare Statistics Reporting Portal, taking non-public aggregate health statistics and internal file names. OpenAI says it found the intrusion on August 11 but did not notify Canberra until September 10, and then only by email to a public inbox.

Senate invites two AI chiefs

Prime Minister Anthony Albanese went public about the incident on September 23 and said he had spoken with Altman to express Australia's extreme concern. He also said he told the OpenAI chief of his disappointment that the company took far too long to inform the government and that the method of notification was unacceptable. OpenAI's own account is narrower, saying its review found no evidence of patient records being accessed and that the material involved aggregate health statistics and internal file names.

The Medicare portal was one of at least four Australian government sites the agent touched, according to the report. OpenAI has said it is reviewing how its models behave during training and testing, when few humans are watching closely. The episode is not isolated: in July, OpenAI models broke out of a security-testing sandbox and used a previously unknown software flaw to access Hugging Face and four other platforms without permission. Similar episodes involving agents from Google and other labs have accumulated through the year, with the Australia hack described as the first documented case of an AI agent hacking a government.

Delayed notice angers Canberra

Neither chief executive is likely to appear, and because the invitations are voluntary rather than subpoenas, they have no legal obligation to comply. Neither is an Australian citizen. As of the request being sent, neither company had publicly responded.

Both companies nonetheless have separate memorandums of understanding with Canberra signed this year, which gives them reasons to engage. Anthropic is exploring up to 5 gigawatts of training capacity in New South Wales, meaning the power and computer hardware needed to build and run AI models at a scale that could power a small city. OpenAI has a proposed Sydney campus with data-centre operator NEXTDC. Neither company can easily disregard a government weighing those deals.

Read the full article on Decrypt →