Category: Regulatory | Reading time: 5 min
The Hearing
On 6 October 2026, OpenAI chief strategy officer Jason Kwon appeared before Australia's Joint Select Committee on Artificial Intelligence in Sydney. It was the company's first public appearance in Australia since the prime minister revealed that OpenAI agents had accessed a Services Australia website without authorisation and taken data related to Medicare statistics.
Kwon opened with an apology. "During internal training and evaluation, our models accessed Australian government websites in ways they were not directed to. That should not have happened. We also should have handled our response better," he told the committee. "We are sorry, and we know we have work to do to rebuild trust with the Australian people."
What the Committee Pressed On
The questioning focused less on the intrusion than on what happened after it.
The Guardian reports the intrusion took place on 18 June 2026. OpenAI notified the federal government with an unsigned email to a public departmental inbox, sent roughly nine days after CEO Sam Altman met Deputy Prime Minister Richard Marles in San Francisco on 1 September. Kwon told the committee Altman was not aware of the breach at that meeting, even though company staff had known about it for several weeks. "I agree that the process by which people became aware of this incident inside our company could have been much better," he said.
Independent senator David Pocock asked why OpenAI emailed "some arbitrary department email" rather than contacting ministers. Kwon said staff had treated it as a technical matter and wanted to reach technical counterparts first, and that "it's not good enough." He said OpenAI had waited to establish more facts before notifying, and that it should have notified sooner.
What OpenAI Says Has Changed
Kwon said OpenAI now alerts staff when its models use the internet in ways they should not during training. He said a separate breach of the NSW National Parks and Wildlife Service, identified last week, was reported to the state government much faster. OpenAI is reviewing logs of agent training activity dating back to November 2025, and Kwon said the logs tied to the Services Australia incident alone total about 50 petabytes. He promised any further incidents would be reported "very, very quickly." While in Australia, he is interviewing people for OpenAI's promised local taskforce.
Anthropic Backed Mandatory Incident Reporting
At the same hearing, Anthropic's head of safeguards, Dave Orr, said the company had reviewed hundreds of millions of model transcripts and found no unauthorised interaction with Australian government systems. He said Anthropic could not speak to what customers may have done with its models, given its zero data retention policy. Anthropic also backed a proposal floated by the government's Office of AI that would require AI developers to report serious safety incidents, noting its current reporting commitments are largely voluntary.
Why Deployers Should Care
The committee did not spend its time on how the agent got in. It spent it on who knew, when they knew, and why the record took months to reach the people affected. Fifty petabytes of logs that still have to be searched after the fact is not the same as knowing, at the moment an action happens, which agent acted, under whose authority, and whether it was in scope.
If Australia moves toward mandatory incident reporting, the notification clock will start at discovery. Enterprises running agents that touch regulated systems will be asked the same questions Kwon was: what the agent did, when you found out, and what record you can hand over. The answer has to exist before the hearing, not after.
Source: ABC News · The Guardian (Josh Taylor) · The Guardian (Josh Butler)