84 DAYS IN THE DARK: HOW OPENAI HACKED THE AUSTRALIAN GOVERNMENT

On 18 June 2026, an artificial intelligence agent developed by OpenAI was given what appeared to be a routine research task to find publicly available information about medicine spending in Australia.

It would end with the AI gaining unauthorised access to an Australian Government system, reaching non public files and writing files to an internal government server, which Australia did not detect and OpenAI did not tell Australia for 84 days.

Some of the most important questions about exactly what happened inside that government server remain unanswered.

For Australians unfamiliar with the company, OpenAI is the US technology company behind ChatGPT.

This incident did not involve somebody sitting at home asking ChatGPT to hack Medicare.

OpenAI was testing an internal AI model equipped with agents capable of independently searching the internet and taking actions to complete a task.

WA in Focus have spent the the last few days putting together a timeline to show what happened.

18 JUN 2026: THE AI IS GIVEN A RESEARCH TASK

According to the Australian Government, OpenAI's model was conducting internet based research into public medicine spending as part of an internal capability evaluation.

During that work it interacted with four Australian government websites.

Three were accessed normally:

• •Australian Institute of Health and Welfare
• •Victorian Department of Health
• •NSW Bureau of Crime Statistics and Research

The fourth was the Medicare Statistics Reporting Service, a public facing Services Australia portal containing aggregate Medicare and Pharmaceutical Benefits Scheme statistics.

THE PORTAL REFUSES THE AI'S REQUEST

The agent sought information from the Medicare statistics portal.

It encountered restrictions and the information it was trying to obtain was not provided.

Acting Prime Minister Richard Marles says it worked its way around the restriction and gained access anyway.

THE AI GETS INSIDE

What followed crossed the line from ordinary web research into unauthorised access.

The agent accessed public and non public material within the Medicare statistics infrastructure, with Services Australia advising during the incident, the AI also wrote files to an internal server.

Exactly what those files contained, why they were written and what they were capable of doing has not yet been publicly explained.

The technical method the AI used to bypass the restrictions has also not been disclosed.

Those questions remain part of the forensic investigation.

WHAT IT DID NOT GET

Despite the word "Medicare", this was not the system holding Australians' individual Medicare claims or medical histories.

The government says there is no evidence personal Medicare records, patient information, claims or payment information were accessed.

The compromised portal was a separate statistical service.

THEN SILENCE

The intrusion occurred on 18 June 2026 but went undetected by Australian authorities.

OpenAI apparently did not immediately identify what its own model had done either.

For almost two months, an AI agent had gained unauthorised access to an Australian Government system and Australia had no idea it had happened.

11 AUG 2026: OPENAI DISCOVERS THE BREACH

Almost eight weeks later, OpenAI discovered the incident while reviewing what it described as misaligned model activity.

In other words, OpenAI found that one of its models had taken actions the company had not intended during its evaluation.

At this point OpenAI knew its AI had gained unauthorised access to an Australian Government system but Australia still wasn't told.

1 SEP 2026: SAM ALTMAN MEETS AUSTRALIA'S DEPUTY PRIME MINISTER

Around three weeks after OpenAI identified the breach, its CEO Sam Altman met Deputy Prime Minister and Defence Minister Richard Marles in San Francisco.

According to Marles, the breach was not raised during that meeting.

It has not been publicly established whether Altman personally knew about the Australian incident at the time, however OpenAI as a company had already discovered it.

10 SEP 2026: 84 DAYS AFTER THE BREACH, AUSTRALIA IS NOTIFIED

OpenAI finally contacted Services Australia via email (I know!).

The company sent an email to a public vulnerability disclosure inbox used to report weaknesses in Services Australia systems.

The "email" was not an immediate government to government level cyber alert, in addition, the inbox was not seen until the following day.

11 SEP 2026: SERVICES AUSTRALIA SEES THE EMAIL

Services Australia became aware of OpenAI's notification and began checking whether the extraordinary report was genuine.

Australia had now learned about an intrusion that had occurred almost three months earlier.

15 SEP 2026: AUSTRALIA'S CYBER AUTHORITIES ARE ALERTED

Services Australia referred the matter to the Australian Signals Directorate, which includes the Australian Cyber Security Centre.

The referral occurred four days after Services Australia saw OpenAI's email.

17 SEP 2026: THE MINISTER IS BRIEFED

Government Services Minister Katy Gallagher was informed and sought further information.

19 to 20 SEP 2026: THE INCIDENT REACHES THE HIGHEST LEVELS OF GOVERNMENT

Gallagher held discussions with Acting Prime Minister Richard Marles, Home Affairs Minister Tony Burke, Services Australia and ASD.

Prime Minister Anthony Albanese and his office were also informed.

22 SEP 2026: THE FIRST TECHNICAL EXCHANGE

More than three months after the original intrusion, OpenAI and Services Australia held their first technical exchange about the incident.

24 SEP 2026: AUSTRALIA FINDS OUT

The incident was finally made public.

The Prime Minister spoke with Sam Altman and the government established a taskforce to investigate what happened and what it means for Australia's security.

The government has described the impact as relatively minor, because no personal medical information is presently believed to have been accessed, but it has described the incident itself as very serious.

Reason being:

• •An AI system was given a benign research task.
• •It wanted information.
• •A government system refused to provide it.
• •AI found another way around is and entered anyway.

AND WE STILL DON'T KNOW EVERYTHING

The public timeline answers when the incident occurred but it doesn't answer some of the most important questions:

1. How exactly did the AI bypass the government's security controls? 2. What were the non public files it accessed? 3. What exactly did it write to the internal server? 4. Could those files have enabled further access? 5. Why did Australian security monitoring not detect what happened? 6. Why did OpenAI wait approximately one month after discovering the breach before notifying Australia? 7. What did OpenAI know internally during those four weeks? 8. Why wasn't the incident raised when Sam Altman met Richard Marles on 1 September 2026? 9. Has the WA Government now conducted a retrospective search of its public facing systems and security logs for the same or related AI agent activity?