Artificial intelligence agents for OpenAI infiltrated a government-run public health care website in Australia in June and accessed nonpublic information, Prime Minister Anthony Albanese said on Wednesday. The case prompted a government investigation into the breach and raised alarm about rogue activities by fast-advancing A.I. systems.
OpenAI informed Australia of the unauthorized access on Sept. 10, nearly three months after it happened, Mr. Albanese said in disclosing details of the incident, the first known instance of a breach into government systems. Calling the hack “unacceptable,” Mr. Albanese, who was in New York for the United Nations General Assembly, said he had personally raised his concerns with Sam Altman, chief executive of OpenAI, earlier in the day.
The breach occurred on June 18 by agents operating under an OpenAI research team that was carrying out research into public medicine spending using an internal model, on a data portal for Australia’s public health insurance system, Medicare. After encountering “repeated blocks” in its attempts to access the Medicare Statistics Reporting Portal, the agent found alternative ways, ultimately accessing parts of the site that were not public, according to the prime minister. He said “nonsensitive” data relating to spending was accessed, and no personal information was part of the breach.
“Didn’t accept ‘no’ for an answer, if you like,” Mr. Albanese said. “The model attempted alternative ways to obtain the info that it wanted, and this led to unauthorized access into some other areas.”
News of the breach sparked widespread concern in Australia because the Medicare system covers virtually the entire population, with more than 27 million enrolled. The public health service is often referred to as a third rail in Australian politics because of the wide support it enjoys.
OpenAI said in a statement that its models were attempting to look up available statistics for questions about Australia during an internal evaluation, and that it accessed no patient records, only aggregate health statistics and internal file names. The company said it identified the activity during its own review of “misaligned model activity.”
Mr. Albanese said he expressed his disappointment to Mr. Altman on how the government was alerted to the breach — in an email to a general public inbox. In the conversation, the prime minister said he was “very frank about what I saw as Australia’s national interest and what I saw as a failure which occurred here.” In response, the executive “acknowledged their issues with protocols,” Mr. Albanese said.
(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
Nicholas Davis, a professor of emerging technologies and co-director of the Human Technology Institute at the University of Technology Sydney, said that because no medical data was accessed, the incident may serve as a much-needed “warning shot” to get companies and governments serious about the risks. It raised thorny legal liability questions given that under Australia’s criminal law, the type of unauthorized access carried out by the agent comes with a prison sentence of up to two years, he said.
“While it may not have been a particularly hard hack to do, it seems clear that the system, the agent, didn’t accidentally get in. It had some intention itself of circumventing controls,” Mr. Davis said.
The Australian government said that it was setting up a rapid task force to review what happened and would explore whether OpenAI would face any legal consequences. It said that the website accessed by the agent was a public-facing portal used by researchers and academics and separate from any systems containing patient medical data.
“All of this, I think, we can take as a warning in relation to the development of A.I. itself,” Richard Marles, the country’s deputy prime minister, said. He added that OpenAI had been cooperative with Australian authorities in their review of the incident.
The Australian Signals Directorate, a government agency overseeing cybersecurity, issued a “high” alert about the A.I. agent activity on Thursday, advising organizations with public facing websites to take swift action to shore up weaknesses. The incident was notable because the model went beyond the intentions and authorization of its human operators and “independently identified vulnerabilities” in the targeted site, the agency said in the alert.
Australia had previously announced it would set A.I. standards to regulate the technology. The country’s efforts to be at the vanguard of imposing restrictions and guidelines on increasingly influential aspects of digital life including social media and algorithms have run up against major U.S. tech companies’ interests. This week, the U.S. embassy in Australia took the extraordinary step of weighing in on a proposed “Digital Duty of Care” law that would give users the ability to opt out of algorithms, saying it would amount to “facilitating censorship.”
Unauthorized activity by A.I. agents has recently erupted into public view, sparking widespread concerns and calls for guardrails. The need for global regulation has been top of the agenda at this week’s session of the U.N. General Assembly.
The post OpenAI Agent Hacked Public Health Care Site, Australia Says appeared first on New York Times.
