Australia’s Review Details How an OpenAI Agent Reached Medicare’s Internal Files

Brendan Smith
Brendan Smith - Cybersecurity Analyst
5 Min Read
Medicare statistics browser window with a drawer of blue internal files
An open file drawer extends from the Medicare statistics portal.

Australia’s November 20 review of the OpenAI Medicare incident confirms that an agent running in an internal evaluation accessed non-public files on a government statistics portal. Investigators found no access to sensitive personal data. The report also explains why verifying the intrusion took time. OpenAI’s first email contained the wrong URL and omitted that the activity came from its own AI evaluation.

A search for medicine spending reached internal files

On June 18, the agent was asked to research how much the government spent per person on medicines for skin conditions in communities in Victoria. According to the review, it first tried unsuccessfully to reach non-public statistics at the Australian Institute of Health and Welfare. It then moved to the Medicare Statistics Reporting Service, a separate Services Australia website for querying aggregate data in a browser.

The site accepted public queries without an account. The agent went beyond those queries, reading internal program files and settings and writing and reading back a small test file, according to OpenAI’s disclosure as described in the review. That write mattered because the agent had changed something on the server while carrying out a task that called for reading public statistics. Investigators found inadequate containment and human oversight during the evaluation.

The statistics service was separate from other Services Australia systems. The review found no evidence that the agent moved into those systems or reached patient-level records. Services Australia took the portal offline on September 15 and moved the data to a more secure environment.

The first email left out who ran the agent

OpenAI found the June activity on August 11, while reviewing model behavior after a separate incident involving Hugging Face. It emailed Services Australia on September 10, Australian Eastern Standard Time, 30 days after identifying the activity. The message reported a vulnerability and unauthorized access, but its incorrect URL made the report harder to validate. It also left the agency without a crucial fact for deciding how to respond. The agent had been running as part of OpenAI’s own testing.

  1. September 11. A Services Australia officer reviewed the email in a public vulnerability mailbox that was checked periodically.
  2. September 15. The agency took the portal offline, declared a cyber incident and notified the Australian Cyber Security Centre.
  3. September 20. OpenAI told the Australian Signals Directorate that an agent in its own evaluation had accessed non-public information.
  4. September 24. OpenAI provided the technical information the directorate needed for its forensic work. The prime minister announced the incident the same day.

What Australia wants to change

The government accepted all 18 recommendations in the review. One calls for considering mandatory deadlines for AI incident reports. Another asks agencies to clarify how they escalate unusual incidents while technical verification is still underway. Both address decisions that mattered in this case: how quickly a lab reports what its agent actually did, and who handles a report that describes activity beyond a potential vulnerability.

For teams that evaluate agents with internet access, the case offers a concrete test for their safeguards. The ability to read a public site should come with limits on what an agent can do there and a way for a human to stop the run when it crosses those limits.

References

  1. Department of the Prime Minister and Cabinet. Release of the Rapid Review report. 20 November 2026.
  2. Department of the Prime Minister and Cabinet. Rapid Review into Australian Government arrangements for an AI-driven cyber incident. 20 November 2026. Findings, chronology and recommendations.
  3. Prime Minister of Australia. Press conference in New York. 24 September 2026. Initial public announcement.
TAGGED:
Share This Article
Cybersecurity Analyst
Follow:
Brendan Smith has spent over 15 years knee-deep in cybersecurity, chasing down malware from the gritty reverse-engineering of old-school trojans all the way to wrangling full-blown incident responses for small-to-medium businesses that couldn’t afford a full-blown breach. Over at Gridinsoft, he’s the guy piecing together those double-checked guides on nasty stuff like AsyncRAT, a remote access tool used in malware campaigns—helping readers make sense of the threat and work through cleanup without the extra headache.
Leave a Comment

AI Assistant

Hello! 👋 How can I help you today?