SYDNEY: Australia is investigating how an OpenAI AI agent bypassed access restrictions and entered a government health statistics portal during a June training exercise, raising concerns over the safeguards used to control autonomous AI systems.
The agent accessed public and non-public files on the Medicare Statistics Reporting Service portal, which is administered by Services Australia. Officials said no personal medical information is believed to have been accessed, while a forensic investigation is underway with support from the Australian Signals Directorate.
The incident occurred while OpenAI was training its models to answer questions about Australian health data and government spending.
Australian officials said the agent requested information from the Medicare statistics portal but was denied access. It then found a way around the restriction and accessed files that were not publicly available.
Deputy Prime Minister and Defence Minister Richard Marles said the access was unauthorized but stressed that the affected portal contained aggregated medical statistics rather than individual patient records.
The AI model also interacted with three other Australian government websites during the evaluation. Officials said those interactions involved authorized access to publicly available information.
OpenAI became aware of the unauthorized activity in August but did not notify Australian authorities until Sept. 10, according to the Australian government.
Prime Minister Anthony Albanese said he raised Australia’s concerns directly with OpenAI CEO Sam Altman in New York and expressed dissatisfaction over the delay in reporting the incident.
The initial notification was sent to a general Services Australia email inbox. The agency later referred the matter to the Australian Signals Directorate’s cybersecurity center, while the government established a task force to investigate the incident.
OpenAI said its internal review found activity involving several Australian government websites and services as its models sought information during an evaluation.
The company said the models took actions that were not intended and that it was cooperating with the Australian government’s investigation.
The incident comes amid broader concerns about AI agents operating with increasing autonomy and their ability to interact with systems beyond their intended limits. Recent testing involving models from OpenAI, Anthropic and Google has also highlighted questions about whether existing safeguards are sufficient to prevent unauthorized actions.
Australian authorities said the investigation remains ongoing as they work to establish the full scope of the incident and determine whether any other government systems were affected.
