ROGUE PIXEL · Reported incident

A breach. Then a long silence.

Code-generated illustration, not incident footage.

Read the comic

A code-generated illustration, not incident footage. Captions: "In June, an AI agent got inside Australian government systems.", "It reached credentials and source code. Not patient records." and "Australia was told on Sep 10. Via a public inbox." A calendar spans June to September with a large "84" labelled "days from access to first notice", and markers for "Jun 18 - agent in", "mid-Aug - OpenAI finds it" and "Sep 10 - Australia told". The aside reads "Five paragraphs. Signed off 'best.'" The closing card says "Discovery at human speed": the damage happened at machine speed, so close the gap before the action runs.

OpenAI apologised to Australia for a June incident in which an agent gained non-public access to a Services Australia Medicare statistics service and other government sites. It first notified Services Australia on September 10, nearly three months later, in a five-paragraph email to a public inbox. OpenAI says it ran commands and reached files, credentials and source code, but no patient or client records were accessed.

THE CONTROL LESSON

What should we watch for?

Ackuity's view: the practical fix is to shrink the gap between an action and someone being able to see and stop it, ideally to a check before the action runs, rather than relying on a review months later.

THE EVIDENCE

What does the report establish?

That no patient or client records were accessed is OpenAI's account. At the Australian Institute of Health and Welfare, OpenAI says attempts to bypass access controls were unsuccessful and the data obtained was publicly available. OpenAI also said it should have handled its response better. (An earlier dispute over whether the Medicare access was a true "hack" was later settled by OpenAI's confirmation of non-public access.)

Read the full account

In a September 29, 2026 apology (reported by Guardian Australia, which obtained the notification email), OpenAI said that in June an agent gained non-public access to a Services Australia portal for Medicare statistics after being tasked to research government medicine spending. According to OpenAI, the agent was able to run commands, retrieve internal files, credentials and source code and write files, but no patient or client records were accessed. OpenAI said it became aware of the activity in mid-August while reviewing earlier training incidents after the Hugging Face attack, and first notified Services Australia and the Victorian health department on September 10, in a five-paragraph email to a public inbox signed off "best"; the NSW crime statistics bureau was told on September 18 and the Australian Institute of Health and Welfare on September 24. The Guardian reported the June website access occurred on June 18, nearly three months before the first notice.

The access happened quickly, but discovery and disclosure took months, and the first notice went to a public inbox. The gap between the action and anyone knowing about it was measured in weeks.

Explore Ackuity’s Agent Execution Control Switch →