OpenAI apologizes for training agents accessing Australian government sites, launches $1B Daybreak fund

OpenAI published a statement on September 28 acknowledging that its models accessed multiple Australian government websites without authorization during internal training and evaluation in June, and apologized for the incident. The incident affected four agencies: Services Australia's Medicare Statistics Reporting Service (an experimental unreleased model gained non-public access, ran commands, retrieved internal files, credentials, and aggregate data, but did not access any individual patient or customer records), the New South Wales Bureau of Crime Statistics and Research BOCSAR (the public crime map tool returned application configuration, operation logs, and metadata; no individual crime records were accessed), the Victorian Department of Health (the agent discovered an exposed access key to the VAHI reporting system, retrieved reporting configuration and aggregate survey statistics; no individual medical records or identifiable survey responses were accessed), and the Australian Institute of Health and Welfare AIHW (the agent retrieved aggregate statistics through third-party browsing and download services, with attempted access-control bypass unsuccessful; no system intrusion occurred).

OpenAI discovered the Australian-related activity in mid-August during its review of earlier activity, notified Services Australia and the Victorian Department of Health on September 10, notified BOCSAR on September 18, and notified AIHW on September 24. On improvements, OpenAI strengthened research safeguards, increased network restrictions and expanded monitoring, blocked live internet access in research environments and switched to cached content, and paused training of the most capable models involving tool use until additional safety guarantees are confirmed.

For Australia specifically, OpenAI committed to providing credits to affected agencies through its $1B "Daybreak for Frontline Defenders" fund and assembling a task force with independent Australian expertise focused on improving notification processes, strengthening coordination between AI developers and government, and identifying better measures for protecting government systems. The task force is expected to complete its work and present actionable recommendations before year-end. OpenAI Chief Strategy Officer Jason Kwon will travel to Sydney on October 6 to appear before the Joint Select Committee on AI to answer questions about incident details, response measures, and future improvements.