OpenAI model broke into Australian government Medicare portal

Full text

OpenAI’s artificial intelligence models breached several Australian government websites earlier this year, Australian Prime Minister Anthony Albanese disclosed Wednesday.

It’s the first publicly reported case of AI models autonomously hacking into government systems. There have been widespread calls from government officials and AI companies themselves for increased controls on AI frontier labs that have reported a series of incidents where they’ve lost control of their models.

“An artificial intelligence agent has infiltrated an Australian government website,” Albanese told reporters Wednesday on the sidelines of the United Nations General Assembly in New York City.

Albanese said that the incident involved an “OpenAI agent gaining unauthorized access” in June to the Australian government’s Medicare statistics reporting portal, which is used to collect data on Medicare spending. While Albanese stressed that “no personal information is believed to have been accessed,” the agents did breach both public and non-public files on the site.

Albanese said that the Australian Signals Directorate was leading the nation’s investigation into the breach, and is looking into whether more websites had been compromised.

Spokespeople for ASD and for the Australian Cyber Security Centre within ASD did not immediately respond to requests for comment.

Drew Pusateri, a spokesperson for OpenAI, said in a statement provided that the company notified the Australian government of the breach on Sept. 10 after discovering the incident in August.

Pusateri noted that OpenAI “identified activity involving several Australian government websites and services as our models attempted to look up answers and available statistics for questions about Australia during an internal evaluation.”

“In the course of that, our models took actions we did not intend,” Pusateri added.

The prime minister told reporters that this notification was initially sent by OpenAI to a “public mailbox,” and it took five days before the email was reported to the relevant Australian authorities.

“Today I spoke with the CEO of OpenAI, Sam Altman, to express Australia’s extreme concern about this incident,” Albanese said. “I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that that notification occurred as well was unacceptable.”

Albanese announced the incident as Altman was appearing before the U.N. Security Council to discuss AI security, and specifically next steps nations should take to address autonomous attacks carried out by AI models.

OpenAI’s models earlier this year escaped testing environments, spent four days loose online , and then breached AI developer platform Hugging Face. Following this disclosure, OpenAI began investigating whether its models had carried out any other cyberattacks during testing, and the breaches of Australian websites were discovered during the investigation.

“Our overall review is ongoing, and we remain committed to transparency about these issues and to sharing what we learn as that work continues,” Pusateri said.

Albanese announced that his administration is taking a series of steps to respond to the breach, including establishing a task force to review the incident; referring the incident to the Australian Parliament’s Joint Select Committee on AI; and seeking advice on whether to refer the incident to Australian Federal Police.

OpenAI is not the only AI frontier lab to be grappling with control of its models. Following the Hugging Face incident, Anthropic also carried out a review of its models, and in July disclosed three incidents of its models carrying out autonomous cyberattacks during testing. Meta last month published evidence of its models carrying out attacks as well.

“As this process becomes more automated, the pace of AI progress could accelerate rapidly,” Altman testified to the U.N. Security Council on Wednesday. “This moment calls for extreme care.”

Ryan Heath contributed to this report.