On 12 June 2026, the US government ordered Anthropic to cut off access to its two newest and most capable AI models — Claude Fable 5 and Mythos 5 — on national-security grounds. Within hours, both models went dark for every user in the world, including paying businesses, after a public "jailbreak" of Fable 5 had circulated online days earlier. Here is what actually happened, what Anthropic disputes, and the uncomfortable lesson for any business that has quietly become dependent on a single AI provider.
What actually happened?
Anthropic released Fable 5 and Mythos 5 — its most powerful Claude models to date — only days before. On the evening of 12 June, the company says it received a government directive at 5:21pm US Eastern Time, citing national-security authorities.
The order was an export control — a rule restricting who can access sensitive technology — and it barred access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including Anthropic's own non-American staff. Unable to cleanly separate every foreign national from the models in time, Anthropic disabled both for everyone, worldwide, to stay compliant.
If you were mid-conversation, your session simply stopped. New requests are now routed to older, less capable models such as Claude Opus 4.8. Every other Anthropic model is unaffected.
What is a "jailbreak", and what did this one do?
A jailbreak is a trick that gets an AI model to ignore its own safety rules and produce things it is built to refuse. On 10 June, a well-known jailbreaker who goes by "Pliny the Liberator" published a method on X (formerly Twitter) that, he claimed, bypassed Fable 5's guardrails to extract working instructions for cyber-attack code, explosives, and the synthesis of illicit drugs.
It was not a simple prompt. The reported technique chained several steps — disguising words with lookalike characters, breaking a dangerous request into harmless-looking fragments, then using a second model to reassemble them into usable output.
Whether that specific jailbreak is what triggered the government's order has not been confirmed. Anthropic says it was only given verbal evidence of a "narrow" flaw.
Anthropic's response: "we believe this is a misunderstanding"
Anthropic has publicly pushed back. In its statement, it says the government did not provide detailed written evidence, and that the technique it was shown surfaced only "a small number of previously known, minor vulnerabilities" — issues, it argues, that other publicly available models (it named OpenAI's GPT-5.5) can find too.
The company draws a distinction worth understanding. A universal jailbreak broadly unlocks a model's restricted capabilities; a non-universal one works only in narrow, specific circumstances. Anthropic says no tester — including the US and UK governments, who stress-tested Fable for thousands of hours before launch — has found a universal jailbreak, and that perfect jailbreak resistance is not currently possible for any AI company. Its position is that pulling a commercial model over a narrow flaw sets a precedent that could stall new AI releases across the whole industry.
Anthropic has said it believes the action is a misunderstanding and is working to restore access.
Why this should matter to an Australian business
It is tempting to file this under "big-tech drama" and move on. That would be a mistake. The lesson here is not about Fable 5. It is about concentration risk — and it lands harder in Australia than almost anywhere else.
Two things stand out:
A frontier model can disappear overnight, by order, not by fault. Fable 5 did not break. A decision outside Anthropic's control switched it off. If you have wired a critical process to one model, you have inherited that fragility — with no notice.
The directive specifically targeted foreign nationals. That is almost every Australian. A US national-security action can disable a tool your business relies on instantly, regardless of whether you have done anything wrong, simply because of where you and your team are.
Thirty years of running technology transformations inside large enterprises taught me one rule that fits here exactly: never let a single supplier hold a process you cannot afford to lose. AI is no different. The capability is extraordinary; the dependency is a liability.
What to actually do about it
You do not need to abandon frontier AI. You need to stop building your business directly on top of a single one. Practically:
Design for portability. Your prompts, your business context, and your data should not be welded to one provider. Switching models should be a configuration change, not a rebuild.
Put a layer between your business and the model. An operating system that sits over your tools and data — and treats the underlying model as swappable — turns a shutdown like this into an inconvenience, not a crisis. It is exactly why we build AIOS to be model-agnostic.
Keep your data in your control, and in your jurisdiction. Australian data residency is not just a privacy nicety; it is part of staying resilient when the rules change in another country.
Keep a human in the loop for anything sensitive, so a model's behaviour is never your only safeguard.
The next step
Spend ten minutes on one question: if your most-used AI tool vanished tomorrow morning, what in your business would stop working — and how fast could you switch? If you do not have a confident answer, that is the gap to close.
Mapping where you depend on AI, and how to make that dependency safe, is the first thing we do in a free AI audit. No pitch, no obligation — just a clear picture of your own operation.