Un agente de IA de OpenAI hackeó la web del servicio de salud de Australia

The Australian Prime Minister, Anthony Albanese, described the incident as unacceptable and discussed it directly with Sam Altman, CEO of OpenAI in New York. OpenAI's artificial intelligence agent illegally accessed a Medicare data portal, the public health system of Australia, according to Albanese, who was attending the UN General Assembly in New York.

This occurred in June when the agent breached the Medicare Statistical Reporting Service portal and, according to Albanese, accessed both public and non-public files. OpenAI explained that the agent took measures that were unforeseen and unauthorized by the company. The AI lab did not detect the problem until August, two months after it occurred, during an internal review of agent behavior initiated after a hacking incident at Hugging Face at the end of July.

One of the points most criticized by the Australian government is OpenAI's delay in communicating the incident. Albanese stated that the company only sent an email to a public mailbox on September 10, and it took five more days for the information to reach the country's cybersecurity agency. The Prime Minister described the situation as unacceptable and met with Altman in New York yesterday to express his displeasure, coinciding with Altman and Dario Amodei (Anthropic) presenting before the UN Security Council on the risks of AI.

For now, both the Australian government and OpenAI agree that there is no evidence of individual citizens' medical information being accessed. Albanese assured that, according to available evidence, there is no broader compromise of the Australian Services Network. OpenAI stated that its review detected activity related to various Australian government websites and services as their models attempted to find answers and statistics available about the country during an internal assessment.

The affected portal publishes data on Australia's public health insurance system, such as billing rates and cost and use of medications. The Minister of Defense and Deputy Prime Minister, Richard Marles, confirmed that the agent accessed multiple systems, including state and federal, including health and crime statistics. Marles indicated that the Government is gathering advice on whether any laws have been violated and is conducting an investigation to determine if other government systems may have been affected.

The revelation came hours after top executives from OpenAI and Anthropic appeared before the UN Security Council, urging greater international cooperation on AI security. Amodei advocated for tight yet achievable international agreements, such as prohibiting the use of AI models for developing biological weapons, and warned that an out-of-control AI could pose a threat to humanity, calling the issue the world's most significant security concern.

However, these proposals for international standards clash with the position of the Trump administration. White House advisor Michael Kratsios rejected any attempt to create a global control scheme for superintelligence during the UN General Assembly, echoing Donald Trump's statements a day earlier at the General Assembly. Amodei and Altman, along with Elon Musk, supported a coordinated slowdown in the development of advanced models and other safety measures, warning of the risks of recursive self-improvement.

Not everyone in the industry shares that caution. Leaders such as Jensen Huang (Nvidia) or Mark Zuckerberg (Meta) argue that security responsibility lies with each company individually. Altman himself acknowledged on Wednesday that even a small risk of AI going out of control catastrophically would not be acceptable and advocated for models to be trained only on the certainty of being able to maintain solid human control.