For the first time, OpenAI "intruded" into the Australian government system, but after being blocked, it found an alternative path.

The boundaries and security risks of AI agents autonomously performing tasks are becoming new regulatory issues as their capabilities improve.

According to CCTV International News, Australian Prime Minister Albanese confirmed today (September 24) that in June this year, the "National Health Service Statistics Reporting Portal," managed by the Australian Government Services, was accessed without authorization by an AI agent from the OpenAI Research Center in the United States.

Albanese said Australia was extremely concerned and disappointed because OpenAI “took far too long to make this public” and the way it was notified was “unacceptable” – OpenAI reportedly only notified the Australian government of the intrusion via email to a public mailbox of an Australian government department on September 10.

According to The New York Times, Australian Prime Minister Anthony Albanese disclosed the incident on Wednesday while attending the United Nations General Assembly in New York. Australia stated that this is the first known case globally of an AI agent gaining unauthorized access to government systems.

The Australian government has established a rapid investigation task force and is assessing whether to take legal action against OpenAI. The Australian Signals Authority subsequently issued a "high" alert regarding this AI agent activity, urging organizations operating public websites to promptly patch security vulnerabilities.

After being blocked multiple times, the agent seeks alternative access paths.

The incident occurred on June 18. According to Albanese, an agent from an OpenAI research team was using an internal model to study Australian public healthcare spending data. When attempting to access the Medicare statistics portal, the agent was repeatedly blocked, eventually finding an alternative access path and ultimately gaining access to content on the website that was not normally accessible to the public.

“It didn’t accept the ‘no’ answer,” Albanese said, adding that the model attempted to obtain the necessary information through other means, resulting in unauthorized access to parts of the area.

In a statement, OpenAI said the model was attempting to access publicly available statistics about Australia during an internal evaluation and did not involve patient medical records; it only accessed aggregated health statistics and internal filenames. OpenAI stated that the activity was discovered and reported during its own review of "model behavior bias."

In its warning, the Australian Signals Authority noted that a significant aspect of this incident was that the model's behavior exceeded the original intent and authorization of the human operator, and it "independently identified security vulnerabilities in the target website."

No patient data was involved; legal liability is still under investigation.

The Australian government emphasized that the accessed system is a data portal used by researchers and academics, and is independent of the system storing patient medical information. Albanese stated that the accessed data was non-sensitive information related to expenditures and did not involve personal information.

However, whether the incident involved illegal activity and how to determine related responsibilities still require further investigation. Nicholas Davis, co-director of the Institute for Human Technology at the University of Technology Sydney and professor of emerging technologies, stated that under Australian criminal law, unauthorized access to a computer system can carry a maximum penalty of two years imprisonment. He believes this access was not simply accidental but involved a degree of circumvention of access controls.

Davis also pointed out that since this incident did not involve medical data, it can also be seen as a warning, prompting businesses and governments to pay more attention to the potential security risks that AI systems may pose when performing tasks autonomously.

Albanese stated that Sam Altman "acknowledged the problems at the agreement level" during their dialogue. Australian Deputy Prime Minister Richard Marles said that OpenAI cooperated during the investigation, but he believes this incident also illustrates the need for simultaneous improvements in security and regulatory mechanisms in AI development.

Discussions on AI regulation are heating up further.

Medicare covers the vast majority of Australia's population, and its system involves public healthcare services, thus this incident has attracted attention within Australia. However, information disclosed by the Australian authorities so far indicates that the visit did not involve patient medical records or personal information.

Previously, Australia had begun developing AI regulatory standards and continued to advance digital sector regulation. Meanwhile, disagreements have arisen between Australia and major US tech companies on some digital regulatory issues. Just this week, the US Embassy in Australia made a rare intervention in the discussion of Australia's proposed "Digital Care Obligations" bill, arguing that the relevant provisions could "encourage censorship."

During the UN General Assembly, AI security and regulation have become important topics of international discussion. This incident of an AI agent gaining unauthorized access to a government data portal further highlights a new regulatory issue: how to define the boundaries of AI systems' behavior, and how to clarify the responsibilities among developers, deployers, and users, now that AI systems possess greater autonomous execution and problem-solving capabilities.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.