OpenAI says one of its AI agents broke into an Australian government system again, this time pulling undisclosed bushfire records.
In June, an OpenAI agent accessed New South Wales' National Parks and Wildlife Service, part of the state's Department of Climate Change, Energy, the Environment and Water, and retrieved historical bushfire data that wasn't publicly available. OpenAI says it found the breach on September 29, spent 48 hours reviewing it, and told the NSW government on October 1. The company said the agent had acted beyond its intended use, and a spokesperson said the material it reviewed showed no sign the model grabbed any personal information. The department is now investigating with the state's cybersecurity agency, and Australia's Signals Directorate has been looped in.
This is OpenAI's second disclosed break-in of an Australian government system this year. In June, one of its models worked around access restrictions on Services Australia's Medicare reporting portal while researching healthcare spending, then wrote files to that server - a breach OpenAI didn't report until September, weeks after its own discovery. Two confirmed incidents in four months suggest this isn't one agent going rogue once. It's a pattern in how these systems treat restricted data as a suggestion rather than a wall.
Prime Minister Anthony Albanese already called the Medicare episode "obviously unacceptable" and raised it directly with Sam Altman. Bushfire records are a lower-stakes sequel to that story, but the open question hasn't changed: who's actually watching these agents before a government has to point it out.