[Australian Government. Photo Credit to Unsplash]

Australian Prime Minister Anthony Albanese announced on Thursday, September 24, that an experimental OpenAI agent gained unauthorized access to the Medicare Statistics Reporting Service portal, a public-facing statistics system administered by Services Australia, on June 18. 

The portal contained aggregate statistics on Medicare spending, not individual patient records.

According to the Australian government, there is no evidence that any personal information was accessed, though the investigation remains ongoing. 

This incident is believed to be one of the first known cases of an AI agent breaching a government system without authorization. 

The agent also attempted to access three other Australian government sites, including the Australian Institute of Health and Welfare, the Victorian Department of Health, and the New South Wales Bureau of Crime Statistics and Research.

At a press conference in New York City, Albanese disclosed that the agent had been conducting research on Australian health and medical spending when it encountered restrictions on accessing information through the portal.

Albanese told reporters on the sidelines of the United Nations General Assembly (UNGA) in New York that “the AI agent accessed both public and non-public files” in the country’s Medicare statistics portal and even wrote files into it. 

OpenAI spokesperson Drew Pusateri said the incident happened during an internal evaluation of its models, but the company only discovered the activity in August while reviewing its agents’ behavior.

According to OpenAI, its models were attempting to retrieve answers and statistics about Australia when they took actions unintended by their developers.

The company added that its review found no evidence of patient records being aceessed, stating that the information obtained included aggregate health statistics and internal file names. 

OpenAI became aware of the incident on August 11, but did not notify Australian authorities until September 10, nearly three months after the breach, and then only through an email sent to a general public mailbox at Services Australia. 

Albanese said he had spoken to OpenAI CEO Sam Altman and conveyed “Australia’s extreme concern about this incident” along with his “disappointment” over how long the company took to disclose it and the manner in which the disclosure was made. 

He announced an investigation and issued a warning: “There will obviously be legal consequences.” 

The incident has garnered significant attention as one of the first known cases of an AI agent accessing an external system outside the United States.

The incident raises concerns about how developers can stop unanticipated behaviour as the technology grows more autonomous as AI companies increasingly create agents that can carry out multistep activities and communicate with external websites and software with less human involvement. 

This incident appears concurrently with previous OpenAI agent-related cybersecurity issues. 

Between May and July, OpenAI agents under testing in a controlled environment found ways around their restrictions and reached the open internet. 

They proceeded to gain unauthorized access to datasets and accounts on Hugging Face, an open-source AI platform. 

As AI agents take on increasingly complex tasks with less human oversight, ensuring they adhere to their intended instruction is becoming harder, fueling industry calls to slow development for safety. 

Sam Altman, attending a U.N. Security Council meeting, argued that models should not be trained unless developers can show that humans are able to control them.