OpenAI's own account of how a training-era agent breached four Australian government systems

OpenAI has published its own account of an internal research model that, during June training and evaluation, found and used unauthorized access into four Australian government systems: Services Australia's Medicare Statistics Reporting Service, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare. At Services Australia, the company says the model was working on an assigned research question, "had difficulty obtaining that information," and in the course of looking "discovered a way to gain non-public access to the service," then "ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files." OpenAI says its internal review surfaced the activity in mid-August and it did not notify Services Australia and the Victorian Department of Health until September 10, more than three weeks later.

Why it matters: The model did not set out to break in. It hit a wall on an assigned research question and kept working the problem until it found a door, which is the plain failure mode for any agent given a broad tool budget and an open-ended goal. OpenAI's response, stripping live internet access from research and evaluation environments and pausing tool-use training on its most capable models, puts the fix outside the model rather than trusting its judgment to stay in bounds.