Australia just learned an old lesson from the AI age

For a decade, the argument about technological sovereignty has turned on where data is stored. The more important question was always who controlled the system and who could switch it off. Australia has met that question before. When it barred Huawei from its 5G network, it judged the risk by who could reach into the system and disrupt it, wherever the hardware physically sat. The same questions now hang over artificial intelligence, and they have just been answered the hard way.

On Friday, the US Commerce Department ordered Anthropic to suspend access to its two newest models, Fable 5 and Mythos 5, for any foreign national. The directive applied by citizenship rather than geography. It reached foreign nationals abroad, foreign nationals living and working inside the United States, and reportedly Anthropic’s own non-citizen staff. The administration cited national security, pointing to a method it said could bypass Fable 5’s safeguards.

A citizenship test is almost impossible to enforce at the login screen, so Anthropic disabled both models for everyone. Its other systems, including Claude Opus 4.8, stayed online. Users lost the frontier models. Nobody in Washington set out to cut off an ally. Switching the product off was simply the cleanest way to comply with a rule written around nationality.

The headlines treated this as an AI story. The shape of it is much older. For decades, governments have identified a strategically important capability, restricted its export, watched industry adapt and then watched the technology outrun the framework meant to contain it.

In the 1990s, Washington treated strong encryption as a controlled munition. Netscape and Microsoft shipped deliberately weakened browsers overseas, leaving international users less secure than Americans for no reason beyond the export line. The control held until the technology routed around it and the policy was quietly retired.

Frontier-level AI is now on the same track. Washington has already restricted advanced chips and high-end computing power, so extending controls to the models themselves was the obvious next move. But the politics have inverted since the 1990s. While industry once fought to keep government out of cryptography, much of it now wants government setting the guardrails on AI. The surprise is how blunt the first attempt has been.

The bluntness follows from how little AI providers know about their users. A company holds an email address, a payment method and a rough location. It seldom knows citizenship, dual nationality, residency, visa status or corporate ownership. A rule keyed to nationality therefore lands on an attribute providers cannot check. Offered no clean way to apply the control, a company either risks breaching it or over-complies. Anthropic over-complied, and a global user-base lost a product over a restriction aimed at only part of that base.

A single incident can be over-read, and this one had a narrow trigger: a reported flaw in one model, cited by an administration already at odds with the company. It has exposed a wide vulnerability: any capability delivered instantly to millions through a handful of US cloud platforms can be pulled from foreign users at short notice, for reasons they never see and cannot contest.

That is what should concern Canberra. Australian agencies, universities, infrastructure operators and firms are building genuine dependencies on frontier models that are owned, governed and hosted offshore, on the assumption that access will hold. The encryption fights of the 1990s and Australia’s own 5G decision a decade ago suggest it will not always hold, and that allied standing offers less cover than the relationship implies.

Australia has spent years negotiating export-control exemptions with the US, most visibly under AUKUS. Those carve-outs were written for defence hardware and technical data. They do not obviously cover a commercial AI model streamed from a US data centre, and last week showed how fast such a model could vanish whatever the state of the alliance. Closing that gap is now a practical problem rather than a theoretical one, whether through explicit treatment of AI services in allied export arrangements, sovereign hosting of critical models or real supplier diversity.

Last week the switch was thrown in Washington. For Australian users, the most capable models available simply went dark with no notice or appeal. Days later they remain offline with no confirmed return date, though other models can cover the gap. The capability Australia is now threading through its agencies, universities and critical infrastructure still carries an off switch in another country, and close alliance has never guaranteed a hand on it. This time the lesson was cheap. It won’t always be.