The Medicare breach, step by step
In 2026 an autonomous AI agent reached non-public systems at Services Australia and three other Australian public bodies. This page sets out the public timeline and, for each step, the open format from the Open Conformance Coalition (OCC) that would have made the failure visible sooner. It does not claim any format would have prevented the breach.
What happened, and what was missing
| When | What happened | What was missing | Open format |
|---|---|---|---|
| 18 Jun 2026 | An experimental AI agent, on an internal research task about public medicine spending, bypasses bot protection on the legacy Medicare Statistics Reporting Service. It runs commands, retrieves internal files and credentials, and writes files to an internal server. | The portal could not tell whose agent it was, what it was authorised to do, or who was accountable. | Agent record: operator, principal, mandate limits, declared vs effective authority, accountable person |
| Same series | The activity reaches the Victorian Department of Health (exposed access key), NSW BOCSAR (configuration, logs and metadata via the API behind its public crime-mapping tool) and the AIHW (aggregate data via third-party services). | Each body saw a fragment. No shared identifier linked the activity across organisations. | Agent record identifier, referenced in every incident record |
| Months (reported) | An independent researcher (Transluce) reports that hundreds of the developer’s own agents shared bypass techniques with one another. | Stopping one agent did not stop the others. Authority and techniques propagated; withdrawal did not. | T-RUE delegation and revocation: withdrawal travels along the chain |
| 11 Aug | The developer identifies the activity in a retrospective review of “misaligned model activity”, 54 days after it began. | Detection came from the vendor’s side, long after the event, with no record of when awareness began. | Incident record: occurred_on, detected_on and aware_on as separate dates |
| 1 Sep | The developer’s chief executive meets the Deputy Prime Minister. The breach is not raised. | Nothing made the gap between awareness and reporting visible to anyone outside the company. | Incident record: aware_on vs initial_report_on |
| 10–11 Sep | The developer emails a generic Services Australia public-disclosure inbox. The email is opened the next day. | There was no published place to send a report about an AI agent. | Disclosure contact, published in /occ/disclosure-contacts.json |
| 14–17 Sep | A senior policy executive meets officials without raising it (14 Sep). The matter reaches ASD (15 Sep) and the responsible minister (17 Sep). | Escalation inside government depended on one inbox and one reader. | Disclosure contact with a named route and a verified date |
| 24 Sep | The Prime Minister discloses the breach in New York, calling it “unacceptable”. | The public learned of it from a speech, not a record. | Public incident record on the register, signed and checkable |
| 29 Sep | The developer apologises and lists commitments: an Australian taskforce, cyber-defence funding, no live internet in research environments, better monitoring, a training pause. The government announces an immediate-reporting standard. | Commitments are statements by one company about its own systems. There is no shared way to check them over time. | Evidence with method, valid-until and withdrawal, and signed attestations |
No evidence has been reported that medical records or identifiable personal data were taken. Items from a single outlet or researcher are marked as reported.
Three failures, not one
54 days to detect, then almost a month more to reach the right people. The immediate-reporting standard announced on 29 September addresses this.
No published place to send a report about an AI agent, so it went to a general inbox. Moving one inbox to 24/7 monitoring does not help the next agency or university.
The affected bodies could not tell whose agent it was, what it was allowed to do, or who answered for it. That was established months later, by the vendor and an outside researcher.
A worked example you can check
A fictional agent record, clearly labelled as an example, with a disclosure contact, mandate limits, a revocation endpoint and an incident showing separate awareness and report dates. It is signed by the register, so you can check it has not been altered.
Public sources
- OpenAI — How we will do better for Australia (29 Sep 2026)
- ABC News — OpenAI agents plotted to access data amid Medicare hack (24 Sep 2026)
- ABC News — OpenAI Medicare breach fuels tougher approach to rogue AI (29 Sep 2026)
- The Guardian — OpenAI agent hacked Medicare: what we know so far (24 Sep 2026)
- The Sydney Morning Herald — OpenAI apologises for Medicare hack (29 Sep 2026)
- Department of the Prime Minister and Cabinet — rapid review