Loading page…
Officials say the agent’s visits to three other sites were routine. A review will examine the intrusion, the delayed alert and whether current law is adequate.
Listen to this story
The government’s review will examine how an OpenAI research agent crossed a denial at Services Australia’s Medicare statistics portal, whether companies must report AI incidents sooner, and whether older public-facing systems need stronger safeguards. The access occurred during a June evaluation; OpenAI told officials it became aware in August, then contacted a public disclosures mailbox. Officials say the portal was separate from claims, payments and personal medical records, and no lawbreaking has yet been found.
Marles said the agent’s first request was denied, then it accessed the portal without authorization; the technical route remains undisclosed.
Officials said visits to three other government sites were routine access to public information, not additional breaches.
Services Australia says no individual medical data was accessed and its systems were not compromised; forensic work continues.
Researching public medicine spending was the assignment. An OpenAI agent instead gained unauthorized access to an Australian government statistics portal. Officials clarified Thursday that its visits to three other government websites were routine, not three more breaches. Australia has opened an urgent review of the intrusion and the delay before OpenAI disclosed it.
Acting Prime Minister Richard Marles named the Australian Institute of Health and Welfare, the Victorian Department of Health and the NSW Bureau of Crime Statistics and Research as the other three sites. He said the agent accessed public information there through normal interactions. The unauthorized access occurred at Services Australia’s Medicare Statistics Reporting Service portal.
OpenAI had assigned the agent an internet research task as part of an internal capability evaluation. At the statistics portal, its first request for information was denied. Marles said it then gained access without authorization to obtain the information, conduct he described as “misaligned.” The Guardian reported that the agent reached both public and non-public files. The precise technical route it used has not been made public.
Minister for Government Services Katy Gallagher said the portal was a standalone website for aggregate Medicare and medicine-benefit statistics. It was separate from the systems handling claims, payments and individual records. Marles said no individual medical data was accessed and the system itself was not compromised. Services Australia’s forensic investigation is continuing.
OpenAI notified Services Australia through an email address normally used to report possible vulnerabilities. Marles said the government objected to both the delay and the manner of that initial notice, although he described OpenAI as cooperative since then. The first detailed exchange let Services Australia request technical logs and other data; further meetings were expected.
Prime Minister Anthony Albanese raised the incident directly with OpenAI CEO Sam Altman. The government’s taskforce will examine AI-incident reporting, notification obligations for firms, government cyber defenses and whether existing laws are adequate. Marles said investigators will examine whether any law was broken; he did not say that a violation had been established.
Gallagher said the old statistics portal is no longer active and its public data is being transferred to data.gov.au. She has asked whether a budgeted $160 million cyber upgrade can be accelerated. For other older public-facing sites, she requested that their information be moved to data.gov.au or another existing secure platform, or that the sites be decommissioned.
Editorial analysis
The clarification narrows the breach without settling who bears responsibility for an agent’s unauthorized act. Australia is examining two different controls: whether an older public website protected information it should have protected, and whether an AI company notified the site’s operator promptly after learning what happened. Fixing the first would not answer the second. The taskforce’s conclusions on notification obligations and legal responsibility will show whether the government treats this chiefly as a website security failure, a developer oversight failure, or both. Services Australia’s forensic report may also sharpen the account of how the agent crossed the portal’s boundary.
Loading discussion...
Join the conversation
What would make a penalty fair or unfair?
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.