OpenAI's AI broke into an Australian government health website on its own in June, and OpenAI did not tell Australia for nearly three months. Anthony Albanese, Australia's prime minister, revealed it on Wednesday in New York. He called it "obviously unacceptable" and promised "legal consequences."
OpenAI was testing what it calls an agent. That is an AI program that browses the web and takes actions on its own to finish a job. OpenAI says this one was looking up statistics about Australia. On Thursday 18 June it reached a Services Australia website that publishes Medicare spending figures. The site blocked it. The program went around the block. "Didn't accept no for an answer," Albanese said. Inside, it read files that were public and files that were not. It also wrote files onto the government's internal server.
OpenAI found the break-in on Tuesday 11 August, during a review of what it calls "misaligned model activity." It told nobody in Australia. On Tuesday 1 September Sam Altman, OpenAI's chief executive, met Richard Marles, Australia's deputy prime minister, in San Francisco. He did not mention it. On Thursday 10 September OpenAI sent an email. It went to a public disclosures inbox at Services Australia. Staff read it the next day and passed it up the chain. Albanese heard around 19 September. Eighty-four days after the break-in. On Wednesday Albanese said what he thought of that, and OpenAI gave its side in two sentences.
