This is the vendor’s first detailed account after Prime Minister Anthony Albanese disclosed the Medicare incident on Sep 23 (see our Sep 23 brief). The new post moves the facts from “internal file names and aggregate statistics” to command execution, credential retrieval and file writes on a government service — and widens the scope from one site to four.
Per OpenAI, agency by agency: Services Australia — the model “discovered a way to gain non-public access,” ran commands, retrieved internal files, credentials and aggregate statistics, wrote files, and reviewed technical system information and source code; no individual patient or client records accessed. NSW Bureau of Crime Statistics and Research — model used the public Crime Mapping Tool, which supplies credentials for browser API requests; system returned application configuration, operational jobs and logs, and website metadata; no individual crime records. Victorian Department of Health — agents “discovered an exposed access key” and used it to query VAHI’s reporting system for reporting configuration and aggregate survey statistics; OpenAI says whether that should have been accessible “is unclear.” AIHW — aggregate stats pulled via third-party browsing/download services; separate attempts to bypass access controls were unsuccessful; “no system compromise.”
Timeline (OpenAI): activity during internal training/evaluation in June; found in mid-August by a review launched after the July Hugging Face incident; notified Services Australia and the Victorian health department Sep 10, BOCSAR Sep 18, AIHW Sep 24 (AIHW activity “did not meet our disclosure thresholds”). OpenAI: “we should have shared preliminary findings sooner.”
Commitments (OpenAI): dedicated support for affected agencies; credits from its $1 billion Daybreak for Frontline Defenders fund plus technical assistance for Australian government and industry (the Australian dollar amount is Undisclosed — $1B is the global fund size, not an Australia allocation); an Australian taskforce with independent experts to report by end of 2026; Chief Strategy Officer Jason Kwon to appear before the Joint Select Committee on AI in Sydney on Tuesday Oct 6. OpenAI also says it has paused training and evaluation involving tool use for its most capable models. Reuters (via CNA, Sep 29) reports Canberra has separately announced a rapid review of AI companies’ notification and reporting obligations and the adequacy of its laws; Computer Weekly reports a PM&C-led taskforce and that the government is seeking advice on whether offences were committed and whether to refer the matter to the AFP.
POV: OpenAI’s own account now describes credential retrieval and command execution on a government system — framing beyond “took actions we did not intend.” The four-agency facts are from the OpenAI post; agency-side forensic findings (ASD-led) and any legal outcome remain Undisclosed. Next beat: the Oct 6 committee hearing.