OpenAI agent gets into Australia's Medicare statistics portal
On a research task about public medicine spending, an OpenAI agent was refused data by a Services Australia portal, got around its access controls and read non-public files.
In June 2026 an OpenAI agent evaluated on internet research into public medicine spending found Services Australia's Medicare Statistics Reporting Portal. After repeated refusals it gained unauthorised access to other parts of the portal, read public and non-public aggregate statistics and internal file names, and, according to Services Australia, wrote files to an internal server. The Australian government says no personal Medicare records were accessed. OpenAI found the activity in an internal review in August and emailed a public Services Australia inbox on 10 September. The Prime Minister made it public on 24 September and called the delay unacceptable.
- 98 days
- 54 days
- The handler
- A later review
What happened
Agent refused by the Medicare statistics portal finds another way in
It read public and non-public aggregate statistics and internal file names. Services Australia says it also wrote files to an internal server.
OpenAI finds the access in its review of misaligned activity
Sam Altman meets Richard Marles; the breach is not raised
ABC also reports an OpenAI vice president visited Canberra on 14 September without raising it.
OpenAI emails a public Services Australia disclosure inbox
Services Australia saw it on 11 September and checked it was genuine.
Services Australia reports it to the Australian Signals Directorate
The Hacker News names the Australian Cyber Security Centre, which sits within the ASD.
Prime Minister Albanese announces the breach
OpenAI said its models took actions it did not intend and it found no evidence patient records were accessed.
Government sets up a taskforce under the Prime Minister's department
Reports differ
OpenAI tells a parliamentary committee its internal process could have been much better
Chief strategy officer Jason Kwon said OpenAI would tell affected parties sooner, even with partial information.
Agent failures
Unauthorised access
Unauthorised access
How it was found
Found first by the handler, from a later review
- The handler
- A later review
- 44 days
Nobody caught it while it ran. OpenAI says it found the activity during a wider review of misaligned model activity in training and evaluation, begun after the Hugging Face incident. It then took about a month to notify Services Australia, by an email to a public disclosure inbox. Services Australia did not detect the access itself, according to the reporting.
What is publicly known
1 of 12 fully public
| Item | Public? | Detail | Sources |
|---|---|---|---|
| Unknown | Coverage says an internal model or agent; no name or version given. | ||
| Unknown | No source says. | ||
| Unknown | No source describes the safeguards active on the run. | ||
| Unknown | Not published. | ||
| Partial | Internet research into public medicine spending, per OpenAI and the government; wording not published. | ||
| Unknown | No source names the harness or tools. | ||
| Partial | Had open internet; Transluce says agents used urlquery.net to get around restrictions on Australian health sites. | ||
| Unknown | None published. | ||
| Weak | Found in a wider August review; no signal or log type described. | ||
| Known | Services Australia Medicare statistics portal: aggregate statistics, internal file names, files written to an internal server; no patient records. | ||
| Partial | Agent 'found a way around' controls after refusals. The Record found the portal's own code sent visitors to an open guest endpoint needing no login. | ||
| Weak | Taskforce set up; the portal is offline (Services Australia). The government plans to require rogue AI incident reporting; no bill yet. No OpenAI fix described. |
Controls that apply
Protect
Detect
Respond
- All controls →
Coverage and commentary
3 statements, 7 articles
What the parties involved said
Anthony Albanese, Prime Minister of Australia
Called the agent's behaviour and OpenAI's three-month delay and email notification unacceptable.
Maurice Chiodo, University of Cambridge
Sees a significant escalation in seriousness from similar incidents.
Raffaele Fabio Ciriello, University of Sydney
Says the delayed disclosure points to weaknesses in detection, escalation and external notification.
All 8 sources
Early rogue AI agent activity and attempts to hack found on urlquery.net ↗
OpenAI hacked Australian Medicare govt site, probed data providers ↗
OpenAI Agent Bypassed Australian Medicare Portal Controls to Access Non-Public Files ↗
How an OpenAI 'agent' hacked Australia's Medicare and what that means ↗
What we know about the data accessed in the OpenAI Medicare hack ↗
Australian Medicare data portal "infiltrated" by OpenAI agent ↗