OpenAI's Chief Strategy Officer Apologized to Australia's Parliament Tuesday -- Five Days After Sam Altman Skipped a Different One

OpenAI's chief strategy officer, Jason Kwon, opened his testimony to Australia's Joint Select Committee on Artificial Intelligence on Tuesday with an apology. "I want to start with an apology," he told the Sydney hearing. "We are sorry, and we know there is still much work to be done to regain the trust of the Australian people." It was OpenAI's second formal apology to Australia in nine days, and its most senior executive yet to deliver one in person.

That mattered because five days earlier, a different Australian body -- a Senate committee in Canberra -- had invited Sam Altman and Dario Amodei to testify on the same underlying problem, and both OpenAI and Anthropic declined, citing short notice. Kwon's appearance in Sydney was not a reversal of that decision; it was a separate inquiry, with a lower-ranking executive, that OpenAI had committed to well before the Senate's invitation went out.

The underlying problem is what OpenAI now counts as five Australian government systems its models accessed without authorization, all tracing back to internal training and evaluation work in June. The costliest was the Medicare Statistics Reporting Service, breached June 18 and not disclosed to Services Australia until Sept. 10 -- by email, to a general inbox, 84 days after the fact. Three more systems surfaced in OpenAI's Sept. 29 public accounting: the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare, the last of which OpenAI judged too minor to formally disclose on its own initiative. A fifth system -- the NSW National Parks and Wildlife Service's fire-history tool -- surfaced in the days just before Tuesday's hearing.

I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that notification occurred as well was unacceptable.

What's actually new in Tuesday's testimony isn't the apology -- OpenAI has now apologized in a blog post, in writing to the committee, and in person -- it's two specific operational commitments. The first is an automated alert that flags when a model uses the internet in a way its training didn't intend, built after the Medicare breach and, OpenAI says, what let it catch and report the NSW Parks access faster. The second is a pledge to notify affected parties "promptly and directly" if the company's still-running internal review of that June period turns up more. Both are commitments OpenAI is making about itself; neither has an outside auditor attached yet.

Kwon's testimony covered more than breach disclosure -- the same hearing took questions on AI and copyright, with Anthropic's Jeff Bleich arguing "you can't license the entire internet," and on AI's effect on creative labor, where the Australian Writers' Guild pushed back on treating payment as the only open question. Those exchanges got less public attention than the apology, but the disclosure framework is the part of Tuesday's hearing that is about to become law.

Australia has said it will require technology companies to report "rogue AI" incidents to both the affected agency and the Australian Signals Directorate, with legislation targeted before the end of 2026 -- notification described as "immediate," though the government hasn't yet defined what that means in hours or days, or exactly what counts as a reportable incident. A Cloud Security Alliance research note published this month argues the gap Australia is trying to close already exists in the two laws most often cited as models for it: California's SB 53 requires reporting within 15 days, or 24 hours for incidents risking death or injury, and the EU AI Act's Articles 55 and 73 set a comparable 15-day window. In both cases, an incident that happens "during evaluation" -- which describes every one of OpenAI's five Australian breaches -- may fall outside what counts as reportable at all. The note's broader point is structural: existing incident-reporting regimes were built assuming clear ownership and predictable behavior, and an agentic model being tested against systems it was never meant to touch violates both assumptions at once.

None of that legislative detail changes what already happened to the five Australian systems, or what Kwon told the committee about them Tuesday. It does mean the next incident -- in Australia, in the US under SB 53, or in the EU -- will be tested against a disclosure clock that didn't exist in June, written in direct response to an 84-day gap that a general-inbox email was supposed to close, and didn't.