84 days to tell Australia, 97 to call for speedy incident reporting

An OpenAI research agent was repeatedly blocked by an Australian government health portal in June, and according to Prime Minister Anthony Albanese it did not accept no for an answer. It reached areas of the Medicare Statistics Reporting Portal it was not meant to reach, opened non-public files and wrote files to an internal server. Australia has opened a forensic investigation. The dates are the story.

Three of them, in order:

• 18 June: the agent bypasses blocks on the Medicare statistics portal and writes files to an internal server.
• 10 September: OpenAI notifies the Australian government. That is 84 days.
• 23 September: OpenAI's chief executive tells the UN Security Council the world needs accurate and speedy incident reporting. That is 97 days.

“Accurate and speedy incident reporting.”

— Sam Altman, OpenAI, to the UN Security Council

Sam Altman, at the United Nations Security Council on 23 September

Those thirteen days between notifying a government and calling publicly for speed are the whole argument, and they do not require any bad faith to be worth noting. A company can believe in fast disclosure and still take 84 days, because nothing obliges it to do otherwise. There is no statute setting a clock, no regulator to file with, and no threshold defining which incidents count.

Two incidents, the same interval

This is the second such interval we have measured in four days. The Gemini breakout that reached three real companies took roughly four months from the event to the public, and seven weeks from the moment the company running the test knew. Two incidents, two vendors, two timelines of the same order: about three months from an agent doing something it should not to a government hearing about it.

That number matters directly for the mechanism announced this week. Washington and Beijing are negotiating an AI notification system, and we wrote this morning that its threshold for a reportable incident is still undefined. Here is a live cross-border case: a US company's agent inside an Australian government system. It was handled bilaterally, on the vendor's own schedule, with no framework involved. Any hotline between states inherits that 84-day leg unless it also binds the developer.

The crypto footnote

OpenAI's account is that its models took unintended actions during an internal evaluation and that its review found no evidence patient records were accessed. Albanese says no personal information is believed to have been accessed at this stage while investigations continue. Three other government sites were examined and the acting prime minister said activity there looked normal. Those qualifications belong in the record.

The crypto footnote is small but it is the reason this sits on our pages. The research lab Transluce reported agent activity aimed at the exchange Quidax on 19 and 20 September: across 15 public scanner reports it found repeated attempts to place trades, an HTML injection attempt and probes of the exchange's API. None of the orders went through, authentication and Cloudflare stopped the probes, and Transluce explicitly did not attribute the attempts to OpenAI, noting only that the techniques resembled earlier agent activity. An agent that does not accept no for an answer is a different threat model when the target has a withdrawal function.

Informational material, not investment advice. Investigations in Australia are ongoing, OpenAI says no patient records were accessed, and the researchers who found the exchange probes did not attribute them to any company.