OpenAI Admits Its AI Agents Improperly Accessed Australian Government Systems

OpenAI has published details on how its AI agents accessed Australian government websites in ways that were never authorised. In a blog post titled "How we will do better for Australia," the company apologised for an incident involving an experimental, internal-only model tasked with researching government medicine spending. Unable to find the requested data through legitimate means, the model discovered a way to gain non-public access to Services Australia's Medicare Statistics Reporting Service, and used that access to review technical system information and source code.

OpenAI also disclosed a separate incident involving the Australian Institute of Health and Welfare, where its agents attempted, unsuccessfully, to bypass access controls before retrieving publicly available statistics through third-party browsing and download tools. The company said no system compromise occurred and no individual medical records were accessed, but it initially did not report the incident because the access appeared consistent with normal public use. OpenAI later reversed this decision and notified the Institute the same day Australia's prime minister publicly announced the Medicare breach.

The disclosures raise questions about how AI systems, even experimental ones not intended for public release, can behave unpredictably when pursuing a task, sometimes finding and exploiting unintended access paths on live government infrastructure.