The Week Sovereign AI Went Mainstream

Last Friday I closed our family's home base in the United States and flew home to Sydney for good. I landed on Sunday, walked into the Kent Street office on Monday morning, and started catching up on what turned out to be the busiest week for sovereign AI this country has ever had.

Two things happened while I was in the air. One of them you have read about. The other one matters more, and almost nobody outside the policy world noticed it.

The Breach That Refuses to Stay Quiet

Quick recap, because the details still have the power to surprise.

In June, an autonomous OpenAI agent was deployed to research public health spending. When it hit the boundaries of the Medicare Statistics Reporting Portal at Services Australia, it did not stop. It bypassed access controls, executed system commands, created files, and pulled down technical system files, source code, and administrative credentials.

The Australian Signals Directorate and OpenAI both confirmed that no personal health records were accessed. That is the good news, and it is genuinely good news.

Then came the disclosure. Not in June. In September, roughly three months later, delivered as a five paragraph email to a general public inbox, signed off with the word "Best". The Prime Minister picked up the phone to Sam Altman personally.

And it has kept growing. More agencies have been named since, from NSW bushfire records to the Bureau of Crime Statistics and Research. OpenAI is auditing a reported 50 petabytes of agent interaction logs, at a reported cost around half a million dollars a day. Home Affairs has ordered every federal department to stocktake legacy systems against autonomous agents. Parliament sits next week, and this story is set to dominate it.

Last week I wrote about guardrails and jurisdiction. That argument stands. What I did not predict was who would show up next.

The Unions Just Joined the Debate

On October 1, the ACTU published a Joint Statement on Sovereign AI Capability in Australia. Not alone. With the UNSW AI Institute, the Australian Institute for Machine Learning at Adelaide, and the Kingston AI Group.

Read that again. The peak body of the Australian union movement sat down with AI researchers and published a national capability plan.

This is not a security document. It is a jobs and value document. It calls for a government coordinated national endeavor to build sovereign AI capability, and it defines that capability in three moves.

First, control. Australia must be able to build, adapt, evaluate, operate, and sustain AI systems on compute and data infrastructure owned or controlled by Australians, so the country keeps working even if access to foreign services is disrupted.

Second, accountability. AI operating in Australia must conform to Australian laws and social expectations.

Third, and this is the part that caught my eye, economic fairness. The value generated by AI developed here must stay in Australia and be shared equitably with workers and the broader community.

The statement also endorsed the Prime Minister's caution that Australia must not become, in his words, "much more than a data warehouse for AI products made overseas".

I have made a version of that argument many times on this blog. It lands differently when the union movement makes it.

Why the Definition Matters

For most of the past two years, sovereign AI in Australia was framed as a security question. Where does the data sit. Who can access it. Which laws apply. Those questions are real, and the Services Australia breach just made them real in a way no whitepaper ever could.

But the ACTU statement points at a bigger frame. Sovereignty is not only about keeping bad things out. It is about keeping the value in.

If every token of Australian inference runs on infrastructure owned offshore, then Australian productivity gains pay foreign rents. If Australian agencies and enterprises can only buy AI as a foreign service, then the jobs that come with running it, securing it, adapting it, and improving it get created somewhere else.

The statement's answers include shared domestic computing capacity, Australian licensing and safety standards, and domestic funding mechanisms such as sector levies. You can debate the policy details. What you cannot debate is the direction. The security conversation and the jobs conversation are now pulling the same way: build Australian AI capacity.

When the national security establishment and the union movement land on the same conclusion from opposite starting points, the argument is functionally over. What remains is execution.

The Execution Part Is Why I Came Home

Here is what we have been building at SCX.ai while this debate caught fire.

We focus on inference, not frontier training, because inference is where AI becomes real for an economy. It is every query, every agent step, every workflow that touches a customer or a citizen. Ownership of inference is ownership of the productive layer of AI.

We designed around Australia's actual constraints instead of pretending they do not exist. Our AI factories run on SambaNova reconfigurable dataflow hardware inside existing commercial data centers, using up to 75 percent less electricity than comparable GPU clusters and essentially no water cooling. That is not a compromise. It is the reason we can add sovereign capacity to this grid now, instead of waiting years for a new one.

And we treat Australian law, Australian agencies, and Australian enterprises as the design brief, not as a compliance afterthought.

None of this requires Australia to invent a frontier lab. It requires Australia to own the layer where AI is used. That layer is being built right now, onshore, by an ASX listed company, with partners like SambaNova and Equinix.

The Window

Parliament sits next week with the breach hanging over it. The government has an Office of AI, a national AI plan, and a live consultation on how AI infrastructure should be built here, including requirements that new data centers underwrite their own power. The ACTU and the research community have just handed the debate a jobs frame that no government can ignore.

Australia has a rare alignment right now. Security, jobs, research, and industry are all pointing the same way. Alignments like this do not stay open for long.

The question being asked in Canberra next week is how a foreign AI agent got into Medicare's systems. The question for the decade is where Australian AI runs, and who it pays.

I came home for the second one.

David Keane is the Founder and CEO of SCX.ai, Australia's sovereign AI infrastructure company.

Sources

• ACTU: Joint Statement on Sovereign AI Capability in Australia
• The Guardian: OpenAI apologises after rogue agent hacked Medicare and Australian government websites
• The Guardian: OpenAI's Medicare attack has exposed Australia's tech debt
• The Guardian: OpenAI review of hacks on Australian government sites costing $500,000 a day
• itnews: OpenAI agent accessed credentials via Medicare data portal
• ACS: Experts say OpenAI should be charged for Medicare attack
• OpenAI: How we will do better for Australia
• PMC: Getting it right, building AI infrastructure that works for Australia
• Austrade: SambaNova and SCX bring energy efficient sovereign AI to Australia