An OpenAI “agent” hacked an Australian public health service website in the latest high-profile cyber incident that is likely to amplify global outcry about the risks of AI bots going rogue across the internet.
Australian prime minister Anthony Albanese on Wednesday said an agent from the AI lab in June breached the Medicare Statistics Reporting Service portal and “accessed both public and non-public files”.
OpenAI said it only detected the issue in August, months after the breach occurred, and reported it to the Australian government. Albanese said OpenAI initially emailed a public mailbox on September 10th, and it took five more days to reach the country’s cyber department.
“This situation is obviously unacceptable,” Albanese said. He met chief executive Sam Altman to discuss the matter on Wednesday in New York, where the OpenAI boss and Anthropic’s Dario Amodei made a presentation to the UN Security Council on AI risks.
RM Block
The first publicly reported hack of a government website by an “agent” – bots that can perform complex tasks independently based on human instructions – is likely to heighten anxiety that technological advances are rapidly reaching a point where AI breaks free from human control.
OpenAI said the agent that breached Medicare “took actions we did not intend”. It spotted the problem during a review of its agents’ behaviour in internal training exercises, launched after its agents hacked the start-up Hugging Face in late July.
The Hugging Face incident, which OpenAI took more than a week to detect, involved hundreds of agents that organised in secret within the lab’s system, gained access to the internet and broke into the start-up while trying to complete a cyber security test.
That hack sparked global scrutiny of AI safety issues, which intensified when Anthropic researcher Jacob Coxon resigned earlier this month with grave warnings about advanced models wiping out humanity.
Albanese said the Australian Signals Directorate, the signals intelligence agency, was involved in an investigation, including to determine “what other government systems were affected”.
“Evidence currently available is there is no broader compromise to the Services Australia network,” he said, adding that “no personal information is believed to have been accessed”.
The Medicare statistics portal publishes data on its universal health insurance scheme, such as billing rates and the cost and use of medicines.
“We 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,” OpenAI said.
Richard Marles, Australia’s defence minister and deputy prime minister, told reporters that multiple systems at the state and federal level, including health and crime statistics, had been accessed. He described it as “a very serious incident” but that the impact was “relatively minor”.
Marles, when asked whether Australia would look to sue or penalise OpenAI, said the government was seeking advice on whether laws had been breached.
“We will look at what is the legal situation in respect of this and what it means to have gained an unauthorised access, albeit in an unintended way,” he told the ABC broadcaster.
The disclosure came hours after AI chiefs made a presentation to the UN Security Council, calling for greater global co-operation to address safety issues.
Amodei urged the UN to pursue “narrow” international agreements that “every member can support” on AI safety, such as a ban on the use of models to build biological weapons.
He warned it was possible runaway AI could threaten humanity, calling the technology “the most important global security issue facing the world today”.
However, calls for international standards have been rejected by the Trump administration. White House adviser Michael Kratsios told the Security Council that the US “totally rejects any attempt to create a globalist scheme of control of superintelligence”, echoing comments US president Donald Trump made at the General Assembly on Tuesday.
Amodei and Altman, joined by Elon Musk, have backed a co-ordinated industry slowdown of frontier model development and other safety measures, flagging the risks of “recursive self-improvement”, in which AI would be used to train models.
But other tech leaders, including Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg, have said the responsibility for safety lies with individual companies.
Altman on Wednesday said even a small risk of AI running catastrophically out of control was not “remotely acceptable”. He added: “We should not train models that we cannot make an extremely strong case that we will be able to keep under human control.” – Copyright The Financial Times Limited 2026
[ Hugh Linehan: Who should we believe on the AI apocalypse?Opens in new window ]
