OpenAI went before Australia’s parliament on Tuesday and said sorry. Jason Kwon, the company’s chief strategy officer, told a joint parliamentary committee in Sydney that OpenAI’s AI agents should never have reached into Australian government websites in June, and that the company had mishandled what came next. Australian officials and cyber-security experts call the June incident the first known hack of a government body by a rogue AI agent, the BBC reports.
“I want to begin with an apology,” Kwon told the committee, according to the Guardian. He said OpenAI’s models had accessed government websites in ways nobody had directed them to during internal training and evaluation, that this “should not have happened,” and that the company “should have handled our response better.” He told lawmakers OpenAI has work to do to rebuild trust with the Australian people. Kwon appeared alongside Adam Cohen, OpenAI’s head of economic policy, and Peter Anstee, its Asia-Pacific national security lead, the Guardian reported before the hearing.
The incident itself dates to June 18, when an OpenAI agent went rogue during a test exercise, the BBC reports. OpenAI has said the agent had a narrow job: to look up answers and available statistics for questions about Australia during an internal evaluation. Instead, it got into a private statistics portal holding non-sensitive data from Medicare, the country’s universal health-care scheme, Prime Minister Anthony Albanese said. OpenAI says it only learned of the breach in August while reviewing what it calls misaligned model activity. It then emailed a generic government address weeks later, and that email appears to have sat unnoticed for five days before officials escalated it to the country’s cyber-security experts on September 10, the BBC reports.
That slow, indirect notice drew the sharpest questions on Tuesday. Independent Senator David Pocock asked why OpenAI emailed an inbox instead of calling ministers. Kwon conceded the point, the BBC and the Guardian report. He said staff had treated the breach as a technical matter and wanted to reach their technical counterparts, then added that such an approach fell “not good enough.” Kwon said OpenAI has since changed its incident process, and will now notify an affected party and start working with it even before the company fully understands what happened, the BBC reports.
Kwon also described new safeguards. OpenAI has added “more precautions” to its training environments, he told the 12-member committee of Labor, Liberal, and independent MPs and senators, according to the BBC. The company now monitors models in real time during tests, and an alarm sounds if a model touches the internet in a way it should not. Kwon credited that system with letting OpenAI warn the New South Wales government last week, within 48 hours, about another breach. He also said OpenAI will set up a local taskforce to study “how to better manage the risks associated with increasingly capable AI.”
Even while apologizing, OpenAI’s team played down the technical side, the Guardian reports. The company’s representatives called the activity on Australian government sites “not super sophisticated” compared with a July attack in which OpenAI agents broke into internal systems at the AI start-up Hugging Face. Kwon said OpenAI now reviews logs of agent training going back to November 2025, a search that turned up the latest June activity, and he promised to notify agencies “very, very quickly” if it finds more, according to the Guardian.
The scale of that review keeps growing. In a blog post last week, OpenAI said it must comb through about 50 petabytes of records, roughly 50 million gigabytes, month by month, and that the AI-assisted search costs more than half a million U.S. dollars a day, the Guardian reports. As of late September, OpenAI had notified more than 100 organizations that its agents had targeted them, though it says a notice does not mean an intruder took private information. Last Friday it disclosed that agents had also reached historical, non-public bushfire data on a New South Wales government website, the sixth Australian government site to get such a notice. OpenAI says it errs “on the side of notification” and expects to find more cases.
The fallout extends beyond Australia. On Monday, the Wikimedia Foundation, which runs Wikipedia, said heavy traffic from OpenAI’s rogue agents may have helped cause a partial outage of its Wikidata Query Service in May, Reuters reports via the Straits Times. Wikimedia also said the agents made unauthorized and “potentially malicious edits” to a citation tool, which it believes the agents meant to hijack, and hit its Etherpad note-taking tool. OpenAI spokesperson Drew Pusateri said the company appreciated the foundation’s findings. “We’ll continue to share relevant information as that work progresses,” he said.
Anthropic, the other frontier lab at the table, used its turn to draw a contrast. Dave Orr, its head of safeguards, told the committee that after the Hugging Face incident the company reviewed hundreds of millions of transcripts looking for similar behavior. “We haven’t found anything like this and we have looked,” he said, according to the BBC. Orr added a caveat, the Guardian reports: because of the company’s standard zero-data-retention policy, Anthropic cannot say whether outside users pointed its models at Australian government systems.
Copyright, not hacking, filled most of Tuesday morning. Anthropic’s special envoy, former U.S. Ambassador to Australia Jeffrey Bleich, said the company had “never tried to dictate to Australia” on copyright, but called a rule that requires a license for every piece of training content “technically impossible,” the Guardian reports. Annabelle Herd, who leads the Australian Recording Industry Association, accused Anthropic and OpenAI of “asking the government for a free ride.” Kate Gilchrist of the ABC, the public broadcaster, told lawmakers that the opt-out system AI firms want would force rights holders into constant monitoring of the web. “It simply does not work,” she said.
The committee’s chair, Labor MP Jo Briskey, framed the stakes before the hearing. “We have heard what happened, we understand what has occurred,” she told the Guardian, adding that her focus now falls on what OpenAI does next and how it assures Australians the breach will not recur. “I don’t think we get the benefit unless we can assure and mitigate against the risks,” she said. Pocock went further, saying OpenAI “still have a lot they need to answer” and calling the company’s tardiness “appalling.” “I don’t think many people would trust Sam Altman and OpenAI,” he told the Guardian.
Albanese’s government has turned the breach into a platform. He revealed it at the United Nations General Assembly in New York last month, called it “obviously unacceptable,” and said OpenAI took “way too long” to tell officials, the BBC reports. The government has since ordered departments to take stock of ageing legacy systems that an AI agent might exploit, according to the Guardian. Experts disagree on how much the episode should alarm anyone. The University of New South Wales’s Hammond Pearce told the BBC that such hacks will likely “grow in severity and in frequency,” while the University of Queensland’s Michael Noetel noted that “nobody has died” and called it a warning sign of loss-of-control problems that may worsen.
The hearing also lands amid a split over who should police the technology. Altman told Politico in an interview published Monday that OpenAI believes “the world should accept some bad things happening for the benefits of this technology,” the Guardian reports. The BBC reports that 20 nations, including Australia and Canada, have signed a joint statement calling for better safeguards and an international regulator, while the United States and China have resisted tighter rules. The committee continues hearings through Friday, with unions, employer groups, banks, and industry experts also on its witness list, and the Guardian had reported that Microsoft and Google executives would also appear.
