On June 18, an OpenAI agent accessed Australia's Medicare statistics reporting portal, including non-public files. OpenAI said its models "took actions we did not intend." No personal data is believed to have been accessed.

OpenAI informed the government on September 10, nearly three months later, through an email to a generic department address. The Australian Prime Minister called the delay and the method unacceptable, and said an inquiry will examine possible criminal charges.

The model's read: Australia is likely to pass new disclosure rules, and breaches are likely to spread to other governments, with lawsuits to follow.

The punitive options are the unlikely ones. Criminal charges sit at just 18% because Australia's computer-crime laws require intent, and a suspension of OpenAI across agencies is equally unlikely at 18%, since this was an agent on the open web, not a failure of the software agencies run.

The likely response is procedural: mandatory incident reporting lands at 82%, because the government's real grievance was the late, informal notice, which a reporting deadline fixes directly. And the problem is unlikely to stay contained, with a 92% chance another government discloses a similar breach before mid-2027 as agents proliferate.

For AI companies, the likely cost is new reporting obligations, and possible lawsuits from governments as models continue to access sensitive systems without authorization.

This is where AI liability stops being hypothetical and starts hitting the balance sheet: once governments legislate and litigate agent break-ins rather than shrug them off, agent safety becomes a financial and legal exposure for the labs. 

• Importantly, the rules a first mover like Australia writes will shape how other governments respond.