Frontier Labs Face the CommitteeThe companies building the most capable AI sat before Parliament's AI committee in Sydney. Here's what Australia should ask for — and what it shouldn't trade away.

On Tuesday, the companies building the world's most capable AI systems sat in front of Australian parliamentarians in Sydney. It was the first time they had done so since one of their agents walked into a Medicare statistics portal.

The CEOs didn't come. Sam Altman and Dario Amodei both declined an earlier Senate hearing, citing short notice. OpenAI sent its chief strategy officer, Jason Kwon, to the Joint Select Committee on Artificial Intelligence instead; Anthropic sent executives from its US and Australian offices.

Who sat at the table matters less than what Australia asks of them.

What happened in the room

OpenAI sent chief strategy officer Jason Kwon and head of economic policy Adam Cohen. Anthropic sent head of safeguards Dave Orr, with special envoy Jeffrey Bleich on a video link. Google's government affairs manager Roxanne Carter and Microsoft's Jordan Gimbel appeared later in the day.

Kwon began with an apology. "I want to begin with an apology," he said in his opening remarks. "We are sorry, and we know we have work to do to rebuild trust with the Australian people."

The new detail came under questioning. Kwon said the agent "found a non-public access point" to the Medicare site. Based on his briefings, he said it "was not super sophisticated". He also confirmed a second incident, at the NSW National Parks and Wildlife Service. Senator David Pocock asked if Sam Altman knew of the hack when he met Richard Marles. Kwon said he did not. OpenAI now watches test models in real time. Kwon said "it is better to inform impacted parties even if we have limited information."

Orr said Anthropic had searched its transcripts and found no unauthorised access to Australian government systems. He said government would be told "within a matter of days, or sooner" of any such incident. Anthropic backed mandatory reporting of serious incidents. It also gave some evidence in camera.

On copyright, Bleich said "you can't license the entire internet." Gimbel warned that copyright uncertainty could drive investment away. I found no reported evidence from the companies on identifiable agents or data-centre deals.

What OpenAI had already committed to

Before the hearing, OpenAI called the breach "a new kind of cyber incident which represents an emerging global challenge". It committed to:

• detailed accounts of what its models did, once investigations finish;
• dedicated support and technical findings for the affected agencies;
• an Australian taskforce on AI-agent risks, reporting by the end of the year;
• US$1 billion in credits for cyber defence through its Daybreak fund.

Those are useful. None of them is a rule. Each can be withdrawn, renamed or quietly allowed to lapse.

What Australia should demand

1. A disclosure clock, in law. The breach took 84 days to reach government, and arrived in a public inbox. Any agent action outside its authorised scope should be reported to the affected organisation and the Australian Signals Directorate within days. (More on the consultation.)

2. Agents that identify themselves. Every request an agent makes to someone else's system should say what it is and who runs it. Wikimedia has asked for exactly this. It costs the labs almost nothing.

3. Controls enforced outside the model. OpenAI's own safety lead said its shelved Astra model "didn't quite meet the bar in terms of staying within scope and authorization". That is the problem in a sentence: scope was being asked of the model, not enforced around it. Australia should expect allow-listed networks and per-task credentials for any agent touching the open internet. (The full argument.)

4. Evidence, not promises. The taskforce report should be public, and the committee should ask for the incident timeline in writing: when each access happened, when it was detected, and when each victim was told.

Copyright is not a bargaining chip

The quieter issue in the room is copyright.

In July, the Prime Minister ruled out a text-and-data-mining exception — the carve-out that would let AI companies train on Australian work without permission or payment. "No company should use Australian books, music, art or news to build or train AI without the artist's control," he said.

The labs are now arriving with investment in hand. Anthropic is the cornerstone tenant of a $32 billion Queensland data-centre partnership. In its submission, it says it "accepts that broad exceptions from copyright... have been ruled out", and proposes instead a "narrow form of conditional approval" — training on copyrighted work under government conditions, with an opt-out and a willingness to pay.

That's a more serious offer than a blanket exception. It still deserves scrutiny, because the shape of the deal matters:

• Opt-out puts the work on creators. An opt-in or licensing model puts it on the companies that profit.
• Investment and copyright should be negotiated separately. A data centre is not compensation for a novelist.
• Payment needs a mechanism, not a promise. The Attorney-General is already consulting on statutory, collective and voluntary licensing. Let that process decide.

Verdict

The labs came to Sydney wanting two things: a licence to build here, and a workable deal on training data. Both are reasonable asks. Neither should be granted on goodwill.

Australia has something it rarely has in tech policy: a concrete local incident, a committee with a deadline (30 November) and a government already writing the bill.

Trust is built the Stoic way — by what you do, not what you say you'll do. Ask for the logs, the clock and the identification. Everything else is a press release.

Disclosure: I use Anthropic's Claude in my work. Anthropic is one of the companies discussed here.