Australia should allow frontier AI firms such as America's Anthropic and OpenAI to train their models locally as "a matter of urgency" so the government can gain "access and influence" over leading AI systems and protect national security, federal departments and agencies say.

The nation's lack of locally trained cutting-edge models is holding back economic productivity and cybersecurity, commonwealth organisations told a parliamentary inquiry on AI in Canberra on Friday.

Abigail Bradshaw, director-general of cyberintelligence body the Australian Signals Directorate (ASD), said while her agency already has access to "the most advanced frontier models available", this access is "entirely reliant on digital connection with the United States".

"That is why we are suggesting additional resilience from the prospect of actually hosting models in Australia," she said.

"... Access to frontier models is a matter of urgency."

Training frontier AI models in Australia would bring "significant benefits for national security", said Bradshaw, who suggested AI companies could be forced to provide government with early access to test the newest models and suggest safety improvements.

ASD is currently "learning largely through media reporting and voluntary conversations about abnormal behaviour or agentic [AI] escapes", Bradshaw said – appearing to reference recent hacking incidents involving AI agents from both OpenAI and Anthropic.

AI models developed by the two US firms "are several leagues ahead" of competitors, she said.

Hosting AI training in Australia could also help "attract and retain Australian technical expertise at the frontier" and help ASD protect Australian infrastructure from attack as the economy becomes more dependent on AI, Bradshaw added.

Helen Wilson, a deputy secretary in the Department of Industry, Science, and Resources, said Australians working in California's Silicon Valley had already been hired to return home and work in Australia's AI Safety Institute.

"One of the benefits of attracting [AI] training here is to bring some of our best and brightest home," she said.

Defence warns against AI 'that's based offshore'

Chris Crozier, chief AI and information officer at the Department of Defence, told the hearing his employer is “very cautious” about using AI tools located in foreign countries, due to risks such as information leaks and the possible impacts of AI 'kill switches' not controlled by Australia.

“We simply can't be working with a technology that's based offshore, where someone can have a bad day and turn it off," he said.

"We need – particularly when we're getting into conflict – to be 100 per cent capable with our military.

“... We want the compute, we want the application, we want the data, we want the decision making all based in Australia."

Cameron Gifford, a deputy secretary in the Department of the Prime Minister and Cabinet, said the government wants to achieve “access and influence” with AI companies which physically locate their AI training in Australia.

This would help law enforcement undertake “much easier discussions” with such firms “so there isn’t a threat to Australians’ lives through misuse of those platforms”, he said.

Copyright the 'key concern' preventing investment

Major AI companies are currently unable to train their models in Australia using Australian content, due to the country's copyright laws.

The federal government is consulting on the issue, with documents tabled last week showing it is considering allowing AI companies to use copyrighted Australian material for training, as it works to attract local infrastructure investments.

While the documents suggested AI companies would be forced to sign financial deals with a substantial number of intellectual property holders, many creatives say Australia's existing copyright laws are adequate and any deals should be signed with all rights holders – not just some.

Jenna Priestley, an assistant secretary in the Attorney-General's Department – which is leading the government's copyright consultations – told the committee Australia's copyright laws are the "key concern" of frontier AI companies.

She said training frontier AI models in Australia would give the government “more control” over compensation conditions, as Australian rights holders currently have “very limited control and payment” from their material being used.

“Part of the policy objective we've been asked to look at is around attracting the investment to Australia to try and give the Australian government more leverage over those conditions," she said.

In its submission to the parliamentary inquiry, OpenAI said, "Australia can play a significant role in the global build-out of AI infrastructure, if the right conditions – particularly copyright policies – and partnerships are in place."

"... Clear, pro-innovation [copyright] policy would put Australia in a competitive position globally, and would act as a market signal to attract investment from developers and infrastructure companies."

"Policy certainty is critical because decisions about where to undertake AI research, model training and related investment are long-term, capital-intensive decisions which are made across competing jurisdictions," the company said.

Any submission to the inquiry by Anthropic is yet to be published at the time of writing.