When the US Pulls the Plug: Australia’s AI Blackout and What It Means for Your Business

It happened on a Friday evening, US time. The Department of Commerce sent a letter to Anthropic at 5:21pm ET on June 12. Within hours, Anthropic had disabled two of its most capable AI models (Fable 5 and the more powerful Mythos 5) for every customer on the planet.

Not just American users. Everyone. Including thousands of Australian businesses that had built workflows, client systems, and daily operations around those tools.

There was no grace period. No wind-down. No “we’ll give you 30 days to migrate.” The order required that any foreign person accessing the models, including Anthropic’s own foreign-national employees, obtain an approved export licence from the Bureau of Industry and Security. Since processing that many licences overnight was impossible, Anthropic killed access entirely. AWS revoked it across all regions at the company’s request.

InnovationAus called it proof that “Australia’s AI fragility is no longer deniable.” They’re right. But what does that actually mean if you run a 12-person accountancy firm in Parramatta, a marketing agency in Fitzroy, or an e-commerce operation in Brisbane?

What triggered it

The official reason was national security. Amazon’s researchers had reportedly found a method to bypass the safety guardrails on Fable 5, a jailbreak that involved asking the model to review a codebase and identify software vulnerabilities. The White House solicited feedback from other AI companies, then moved.

Anthropic disputed the finding. The company said the jailbreak didn’t work the way the government claimed, and that “if this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”

That dispute didn’t change anything. The access was still gone. Australian businesses woke up to broken workflows and no timeline for resolution. The legal mechanism was the US Export Administration Regulations, specifically the EAR controls on advanced AI models, which give the Bureau of Industry and Security broad authority to restrict foreign access with very little notice.

Why this is an Australian business problem

Most Australian SMBs using AI tools have built their workflows around one or two providers. ChatGPT for content. Claude for drafting and analysis. Maybe one other tool for a specific task. That concentration made sense when these tools were new and there were only a handful of options. It looks different now.

When Fable 5 and Mythos 5 went dark, businesses that had built client-facing systems on those models had no fallback. Not because alternatives don’t exist, but because they hadn’t tested or integrated them. Switching under pressure, with no notice, is expensive and slow. The businesses that fared best were the ones that already had two providers in their stack and kept their prompts and workflows documented independently of any single tool.

Australia sits at the end of a very long supply chain for frontier AI. We don’t produce these models. We don’t regulate them. And as this episode showed, we don’t get a vote when a US government agency decides access needs to stop.

Three things to do about it

Diversify your AI stack. If every AI-assisted task in your business runs through one provider, you have a single point of failure. Pick a second provider and actually use it for something, even if it’s just a secondary option for content drafts or customer communication. OpenAI, Google Gemini, Mistral, and local model options all exist. Know what you’d switch to if your primary tool went dark tonight.

Document your prompts and workflows independently. Your prompts are intellectual property. If they live only inside a provider’s interface, you lose them the moment access is cut. Keep a local copy of every prompt that’s doing real work in your business, along with the context and instructions that make them work. A simple document or folder is enough. This costs you nothing and takes an afternoon.

Read your vendor’s terms on access termination. Most AI provider terms include force majeure clauses that cover government orders. That means if access is cut due to export controls, the provider has no liability to you. Know this before you build a client-facing system on top of any single tool. Plan for the possibility that access ends without warning, because that’s exactly what happened here.

What Australia is doing about it

The federal government’s AI policy push has included calls for domestic AI infrastructure investment, with the Albanese government pointing to sovereign capability as a priority in its AI policy framework. The argument is that Australia needs models trained locally, on Australian data, with Australian compute, so that access can’t be switched off by a decision made in Washington.

That’s a reasonable long-term goal. But local sovereign models at the frontier level are years away, and small businesses need to operate in the world as it is now. The practical answer in the short term is the same as it’s always been for supply chain risk: don’t depend entirely on one source you don’t control.

The Fable 5 and Mythos 5 shutdown was the clearest possible demonstration of what that risk looks like in practice. It won’t be the last one.

More hot takes: Hot Takes →

Sources

• InnovationAus: Australia’s AI fragility is no longer deniable
• Anthropic public statement on model access restrictions
• US Bureau of Industry and Security: Export Administration Regulations (EAR)

📖 More deep dives: News Deep Dives →

📖 More deep dives: News Deep Dives