Kwon tells Sydney inquiry the access should not have happened

OpenAI chief strategy officer Jason Kwon appeared on 6 October before the Joint Select Committee on Artificial Intelligence in Sydney. The Guardian quotes him: “I want to begin 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.” It also quotes: “We are sorry, and we know we have work to do to rebuild trust with the Australian people. We are committed to doing that work.”

The BBC drops the comma and the second sentence, and it reports the same two clauses. This follows OpenAI’s 28 September account and the prime minister’s comments.

Earlier notice, including to ministers

Asked why OpenAI had not contacted ministers directly, Kwon called that a mistake. The BBC quotes: “In retrospect, we should have done what you’re suggesting,” and: “The reason why it happened the way that it did is I think people were thinking about this as a technical situation and they wanted to contact the technical counterparties, but it’s not good enough.” news.com.au, on questions from Senator David Pocock, has “what you were suggesting.”

The BBC then quotes the new practice: “Even if we don’t fully understand the situation, we are just going to notify and start working through the situation collaboratively with the impacted party.” Information Age quotes him calling the Services Australia access “not super-sophisticated” and saying OpenAI “should have informed the impacted parties much sooner.” news.com.au quotes him saying Sam Altman was “not aware at the time” of a 1 September meeting with Deputy Prime Minister Richard Marles.

The NSW line is the parks case already disclosed

The BBC says training runs are monitored in real time, with an alarm if a model reaches the internet in a way it was not meant to. It says that let OpenAI alert New South Wales to “another hack last week” within 48 hours. It does not name the agency.

news.com.au says the NSW Parks and Wildlife Service was notified within 48 hours. The Guardian calls it a separate incident at the NSW National Parks and Wildlife Service, revealed last week, with the activity in June. That is the service in our 5 October article, which quoted OpenAI’s page: identified on 29 September, initial contact within 48 hours. Today’s reports match that disclosure. They do not add a new named incident. OpenAI’s page did not load for this article, so 29 September is the date from that earlier report.

The BBC says Kwon supports a mandatory disclosure framework because it would set “clear expectations.” It also says a local taskforce will look at “how to better manage the risks associated with increasingly capable AI,” the commitment already in the 5 October article. ABC News says he will interview candidates for it while he is in Australia.

Anthropic safeguards lead Dave Orr, in the BBC, said the company reviewed “hundreds of millions of transcripts” after the July Hugging Face case. “We haven’t found anything like this and we have looked,” he said.

Program up, transcript not

The hearings page lists 6 and 7 October in Sydney and 8 and 9 October in East Melbourne, through Friday, as the BBC said. The Tuesday program puts OpenAI at 2.00 pm in the Macquarie Room at NSW Parliament. The 18 September Canberra hearing has transcript links. The 6 October row does not.

Practical takeaway

These quotes are news reports. The apology matches across the Guardian and the BBC. The BBC’s unnamed NSW notice is the parks case already disclosed.