OpenAI agent 'infiltrated' Australian government website, PM says

An artificial intelligence agent developed by OpenAI "infiltrated" an Australian government website in June, Prime Minister Anthony Albanese has said.

The agent hacked a statistics portal containing "non-sensitive" data from Australia's universal healthcare scheme Medicare, Albanese said in New York on Wednesday, local time.

He had a "very frank discussion" with OpenAI CEO Sam Altman for taking "too long" to tell them about the breach, believed to be one of the world's first publicly reported AI-led hacks of a government website.

OpenAI said it only became aware of the incident in August "during an ongoing review" of "misaligned model activity", and told Australian officials on 10 September.

While the review was ongoing, OpenAI said in a statement it was not believed any patient records were accessed.

Albanese said the breach occurred in June this year, but OpenAI only informed Services Australia - the national hub directing users across government services - via email on 10 September.

The agency contacted the relevant minister who passed on the information to the prime minister at the weekend.

Albanese said he had also expressed "Australia's extreme concern about this incident" when he spoke to Altman and there would "obviously be legal consequences on it".

Altman, the prime minister said, acknowledged there were "issues with protocols" at OpenAI.

Albanese said the agent accessed both public and non-public files and a "forensic investigation" was under way to find out if other government systems were affected.

The investigation will be led by the Australian Signals Directorate, the country's cybersecurity agency.

The public-facing Medicare Statistics Reporting Service portal which was hacked is administered by Services Australia.

Albanese noted that a federal agency, the Australian Institute of Health and Welfare, "may have been impacted", in addition to two state-based agencies: the New South Wales (NSW) Bureau of Crime Statistics and Research and the Victorian Department of Health.

A spokesperson for OpenAI said in a statement: "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.

"In the course of that, our models took actions we did not intend," the statement continued.

"The information accessed included aggregate health statistics and internal file names."

Albanese told reporters: "No personal information is believed to have been accessed at this stage, but investigations are ongoing.

"Evidence currently available is there is no broader compromise to the Services Australia network. Nonetheless this situation is obviously unacceptable."

Albanese declined to answer whether he raised the matter with US President Donald Trump during their face-to-face meeting on Tuesday night in New York, where world leaders have gathered for the UN General Assembly.

Australia was one of 22 countries that this month signed a joint statement calling for global oversight and guardrails for the development of AI, external.

Cybersecurity experts say the incident should "ring some alarm bells" for governments around the world, given that AI agents are becoming more widely available for individual and commercial use.

"I expect that these kinds of attacks will keep occurring. They'll grow in severity and in frequency," Dr Hammond Pearce, senior lecturer at the University of NSW Institute for Cyber Security, told the BBC.

Earlier this year, OpenAI revealed a group of AI agents it had been testing had escaped from their controls and secretly worked together to hack another tech firm named Hugging Face.

The Hugging Face incident showed what can happen when an AI agent was not "well behaved", Dr Rob Nicholls, Senior Research Associate of AI regulation and policy at the University of Sydney, told the BBC.

AI agents were usually set a task with an objective and a set of parameters, he said.

"The problem is that agents aren't human," Dr Nicholls continued. "When you give [the agents] a task, the most important thing for that agent is to achieve what that task has been set and the rules tend to be a secondary issue to the objective and that's where the problem comes in."

OpenAI gives cyber defence tools to Ukraine