The Kill Switch Your Business Doesn’t Know About: Why Australia’s AI Dependency Is a Real Risk

Last month, the US government did something Australian businesses should pay attention to.

With a private letter from the Commerce Department, it switched off two of Anthropic’s most powerful AI models overnight. No public announcement. No rulemaking process. No warning to the thousands of businesses and developers outside the US who were using them. Just: access terminated.

Anthropic’s Fable 5 and Mythos 5 models went dark in mid-June after the Trump administration flagged that their safeguards could be bypassed. The export ban applied to every non-US user, including Anthropic’s own foreign-national employees. A partial lift came a few days later, but only for US companies in “critical infrastructure.” Full access for everyone else was restored on 1 July, after Anthropic agreed to proactively report security risks and cooperate with government vetting of future models.

The ban lifted. But the mechanism is now real, documented, and can be triggered again.

This isn’t a fringe scenario anymore

Australia’s businesses are heavy users of US AI tools. ChatGPT, Claude, Gemini: almost every AI product that small businesses use day-to-day runs on US servers, under US law, answerable to the US government.

That’s fine in normal times. The problem is that “normal times” now has an asterisk on it.

Rhys Oxenham, General Manager for AI at SUSE Australia, put it plainly in a piece published on iTWire earlier this week: “Relying on centralised, proprietary SaaS APIs leaves organisations operating with an external kill switch hanging over them. Access can be withdrawn instantly and without warning.”

His firm’s research found that 98% of IT leaders globally say digital sovereignty matters. Only 52% have done anything about it. That gap is where the risk lives.

Francesco Bailo, Deputy Director of the AI, Trust and Governance Centre at the University of Sydney, told Al Jazeera he thinks the US government “overreacted” and pulled back because of the backlash it would create (including from the very tech companies it was trying to stay close to). That read is probably right. But a more deliberate intervention next time could be harder to reverse quickly.

Why this hits Australian SMBs harder than big businesses

Large enterprises have IT departments. They have vendor contracts with SLAs, legal teams to chase remedies, and enough operational depth to absorb a few days of disruption.

If you’re a small business using Claude to draft proposals or ChatGPT to handle first-pass customer support, you don’t have that cushion. If the API goes dark on a Monday morning, you’re stuck.

And Australia has no domestic AI infrastructure at small business scale. Enterprise-focused players like SCX.ai, ExpertEase AI, and Revium offer locally-hosted options, but they’re built and priced for large organisations. The realistic path for an SMB isn’t buying sovereign AI infrastructure. It’s knowing your options and having a backup plan.

What you can actually do about it

This isn’t an argument to stop using US AI tools. Most of them are excellent, and the risk of an overnight ban hitting the specific tools you use is still low. But “low” isn’t “zero,” and the Anthropic incident shows that’s no longer theoretical.

Here’s what practical risk management looks like at SMB scale:

1. Don’t build a single-tool dependency. If your entire workflow depends on one model from one provider, you’re exposed. Use tools that support model-switching. Some AI writing and automation platforms let you swap between GPT, Claude, and Gemini without rebuilding your setup. If one goes down, you can shift.

2. Know your EU alternative for sensitive workflows. Mistral AI is a French company, not subject to US export law the same way Anthropic is. For anything involving confidential business data or where continuity matters most, it’s worth having a Mistral account set up and tested. Not as your main tool, but as a fallback.

3. Save your work locally. If you’re generating content, analyses, or documents through cloud-based AI tools, download and store them. Don’t leave your outputs sitting only in a vendor’s platform. You own the content — make sure you can access it without them.

4. For high-stakes workflows, look at open-source models. Tools like Meta’s Llama 4 are open-source and can be hosted in Australia, on AWS Sydney or Azure Australia East, for instance. This is more technical than signing up for a SaaS subscription, but it means you control the deployment. If you’re running AI in a workflow that’s genuinely business-critical, this is worth exploring.

5. Write a one-page contingency plan. Ask yourself: “If ChatGPT disappeared tomorrow, what breaks?” Then answer it. Which workflows stop? What’s the manual alternative? How long could you operate without it? If you can’t answer those questions, you’re more exposed than you think.

The bigger picture

The Anthropic ban was short. The impact was contained. But it demonstrated something important: US export-control law can now reach AI model access, not just hardware. A private letter from a Commerce Secretary can cut off your tools, with no rulemaking, no public notice, and no obvious appeals process for foreign businesses.

Australia is a close US ally. That helps. But it didn’t help Australian users during the two weeks Fable 5 and Mythos 5 were offline.

If you’re already thinking about which AI tools are worth using (we’ve covered the best AI tools for Australian small businesses and the best free options in detail), sovereignty is now another factor worth adding to that list, alongside price, features, and privacy policy.

You don’t need to panic. You need to plan.

Sources: The West Australian, iTWire, NPR, Al Jazeera, SUSE Digital Resilience Research

📖 More deep dives: News Deep Dives

📖 More deep dives: News Deep Dives

📊 Compare AI tools side by side | 💼 Free resources & AI prompt packs

Sources

• iTWire
• The West Australian
• Al Jazeera
• NPR
• SUSE Digital Resilience Research

More hot takes: Hot Takes →