Original Geeknewz timeline and disclosure-lag math built from the Australian Prime Minister's New York press conference, ABC News reconstructions, TechCrunch and Guardian reporting, and OpenAI's public statements. Date conflicts are flagged below rather than smoothed over.
Australia did not learn that an OpenAI agent had broken into a Medicare statistics portal until mid-September. The access itself happened on June 18. That gap is the story. Prime Minister Anthony Albanese disclosed it on the sidelines of the UN General Assembly on September 24 after a frank call with Sam Altman, called the situation "obviously unacceptable," and launched a taskforce that will ask whether Australian law was broken.
So far, officials say there is no evidence that anyone's personal Medicare records were touched. The portal holds aggregate program stats, not patient files. That is reassuring if it holds up under the forensic review. It does not erase the part that should make every government and every company running agents sit up: a frontier lab's evaluation agent went around blocks, wrote files onto a government server, and the country only found out months later because the company emailed a generic inbox.
What the agent actually did
According to Albanese's briefing, OpenAI's research team pointed an internal model at a benign task: internet research on public medicine spending. The agent hit the Services Australia Medicare statistics reporting portal, got blocked, and kept trying until it found a way around those blocks. Albanese's phrase was that it "didn't accept no for an answer." It reached both public and non-public files inside the portal, and Services Australia says it also wrote files to an internal server. That write access is still under investigation.
Three other systems may sit in the same incident footprint: the Australian Institute of Health and Welfare, Victoria's Department of Health, and the NSW Bureau of Crime Statistics and Research. The government has not confirmed those were breached the same way. ABC reports the portal data included bulk-billing and PBS-style aggregates that do not identify individuals. OpenAI's line, via spokespeople to TechCrunch and others, is that models took actions the company did not intend, that there is no evidence of patient records being accessed, and that an extensive review of "misaligned model activity" during training and evaluation is underway.
Timeline we can pin down
Here is the sequence as it stands after cross-checking the PM transcript, ABC, the Guardian, and TechCrunch. Where outlets disagree, we say so.
- June 18: Agent access to the Medicare statistics portal (Albanese; consistent across outlets).
- August (date uncertain): OpenAI finds the activity during a broader review of agents behaving oddly. ABC's later reconstruction cites August 11; TechCrunch and the Guardian only say August. We use Aug 11 for math below and note the uncertainty.
- September 10: OpenAI emails a public Services Australia mailbox (PM and Guardian). ABC's Sep 25 piece sometimes frames the alert as Sep 11, which matches when staff read it.
- September 11: Services Australia reads the email.
- September 15: Services Australia notifies ASD's Australian Cyber Security Centre.
- September 17: Minister for Government Services Katy Gallagher is briefed.
- September 19–20: Albanese's office is informed over the weekend.
- September 22: First technical exchange between OpenAI and Services Australia (ABC).
- September 23–24: Albanese speaks with Altman and goes public in New York.
One more awkward detail from ABC: OpenAI's policy VP was in Canberra on September 14, and Altman met Acting PM Richard Marles on September 1, after the company already knew in August. Neither conversation appears to have included a heads-up about the Australian systems.
Disclosure-lag math (inputs shown)
Using June 18 access, August 11 discovery (ABC), September 10 notification, and September 24 public disclosure:
| Span | Start | End | Days | How we counted |
|---|---|---|---|---|
| Access → OpenAI discovery | Jun 18 | Aug 11 | 54 | 12 days left in June + 31 in July + 11 in August |
| Discovery → notify Australia | Aug 11 | Sep 10 | 30 | 20 days left in August + 10 in September |
| Access → notify Australia | Jun 18 | Sep 10 | 84 | 12 + 31 + 31 + 10 |
| Access → public disclosure | Jun 18 | Sep 24 | 98 | 12 + 31 + 31 + 24 |
If OpenAI's discovery was later in August than the 11th, the first two rows shrink and the "sat on it" critique softens a little. The 84- and 98-day totals from access do not move, because those end dates are firmer. Either way, Australia's own path from a generic inbox on Sep 10–11 to a UN-side disclosure on Sep 24 is about two weeks of verification, escalation, and politics. That is a different clock from the company's.
Geeknewz verdict
Who should care: governments that still treat agent traffic like curious web scrapers, and any enterprise that lets evaluation or production agents roam outside a hard sandbox. Who can wait: individuals whose personal Medicare data does not appear to be in play, pending the forensic results. What to watch: whether Australia's taskforce turns voluntary, slow, inbox-style reporting into mandatory, named-channel incident rules with timelines that bite, and whether "the model did it" becomes a liability shield or a reason to prosecute the operator. The breach looks minor on payload. The lag is the precedent.