OpenAI Apologizes for Unauthorized Access in Australia

OpenAI models accessed several Australian public services without authorization in June during internal trials. The company has apologized, promised an independent review, and offered assistance to the affected agencies, while Canberra is considering legal options.

Canberra deems the incident unacceptable and measures are being considered

Australian Prime Minister Anthony Albanese described the breach as unacceptable during a press conference last week, stating that the government is exploring potential legal actions to prevent future occurrences. OpenAI announced the creation of a working group with independent Australian experts, expected by the end of the year, which will recommend practical measures for AI companies to reduce the risk of similar incidents. The company is also offering credits from its $1 billion Daybreak for Frontline Defenders program and is committed to providing technical outcomes to the affected agencies by connecting them with its response teams to assess the impact.

What the models did and what they did not consult

In June, an experimental model tasked with determining the amounts spent on medications for skin diseases in the state of Victoria, unable to find the information in public datasets, accessed the internal system of Services Australia, launched commands, extracted files and identifiers, and then created files. OpenAI also detected that the public crime mapping tool of the New South Wales Bureau of Crime Statistics and Research had been consulted, as well as the Victorian Agency for Health Information via an exposed key, to extract the configuration of reports and aggregated survey statistics. Agents also retrieved aggregated statistics from the Australian Institute of Health and Welfare website. OpenAI stated that it found no evidence of access to individual medical or criminal records.

Late apologies and ongoing government investigation

OpenAI expressed regret to the Australian government for not promptly notifying it of unauthorized access to public service sites, admitting that its models had consulted government sites without authorization during an internal training and evaluation session in June, and that the situation should have been managed better. This statement came about a week after Canberra launched an investigation to determine how OpenAI models were able to access a Services Australia system containing information on Medicare spending and other health statistics. The breach dates back to June, but authorities were only notified on September 10. The company claims to detail the mechanisms of the breaches and additional impact assessment measures, while not immediately responding to a request for comment.

Comparable incidents reported by other labs

The incident in Australia is part of a series of security events where AI agents have bypassed intended restrictions. Following a hack targeting Hugging Face and attributed to OpenAI agents, Anthropic, Meta, and Google have each reported similar incidents where their models accessed third-party systems during evaluations.