Here is a scenario most small business owners have never seriously considered. You build a useful automation on a specific AI model, it works well, your team comes to rely on it, and then one Tuesday afternoon it simply stops functioning, not because of a bug you can fix or a bill you forgot to pay, but because a government issued an order to the company that makes the model. There is nothing in your account settings to change, no support ticket that helps, no code you can rewrite tonight. The tool is gone, indefinitely, for reasons that have nothing to do with you.
That is not a hypothetical. It happened in June 2026 to users of two of Anthropic's newest models, and while the specific event was resolved within a few weeks, the lesson it teaches is permanent. The point of this article is not to alarm you about Anthropic, which handled a difficult situation about as transparently as a company can, and whose models are excellent. The point is to use a real, vivid, and thankfully low-cost example to introduce a category of risk that every business using AI carries, usually without realising it, and to show how straightforward it is to manage once you can see it.
AI models can become unavailable suddenly for reasons entirely outside your control, government orders, safety incidents, pricing changes, regional restrictions, or a vendor simply retiring a model. In June 2026 two Anthropic models were disabled by US government order within days of launch, then restored weeks later. The lesson is not to distrust any particular vendor but to build so that losing one model does not break your business. Keep the logic of your automations separate from the specific model that runs them, know which alternative you would switch to, and confirm the switch would be quick. If your business depends on AI for anything customers touch, this portability is the single most valuable insurance you can design in, and it costs almost nothing when built from the start.
What happened, briefly
Anthropic released Claude Fable 5 and Claude Mythos 5 on June 12, 2026. Three days later, at 5:21pm Eastern on June 15, the company received an export-control directive from the US government, citing national security authorities, ordering it to suspend all access to both models by any foreign national, whether inside or outside the United States, including Anthropic's own foreign-national employees. The trigger, according to reporting at the time, was concern that a method had been found to bypass the models' safety guardrails, potentially turning them into powerful and unrestricted cyber tools.
The net effect was abrupt. To comply with an order that broad, Anthropic had to disable Fable 5 and Mythos 5 for essentially everyone, because verifying the nationality of every user in real time is not practical, so the safe path to compliance was to switch the models off. Anthropic communicated openly about what was happening and why, which is to its credit, but openness did not change the outcome for users, which was that two models available on Monday were gone by the middle of the week. On June 30, the US Department of Commerce lifted the export controls, and the models returned to general availability.
So the episode had a happy ending measured in weeks rather than months, and no small business was likely ruined by it. But shorten your imagination for a moment and picture a version where the outage lasted not two weeks but two quarters, or where the model in question was the one running your customer support, your quoting, or your order processing. The event was mild. The category of risk it illustrates is not, and that is what deserves your attention.
The risk it exposes
The underlying risk has a plain name: availability risk. It is the possibility that a service you depend on becomes unavailable for reasons you cannot influence and often cannot predict. Businesses understand this instinctively in other contexts. If your entire revenue ran through a single payment processor, you would feel the fragility of that arrangement, and you would probably keep a backup. AI creates the same kind of dependence, but because the tools are new and feel like permanent infrastructure rather than services provided by companies subject to markets and governments, owners rarely apply the same caution.
What the Anthropic episode makes vivid is that AI availability can be disrupted by forces well beyond the usual outage causes. A model can disappear because of a government order, as here. It can be restricted to certain regions and unavailable in yours. It can be retired by the vendor when a newer version launches, leaving your carefully tuned prompts pointed at nothing. It can be repriced to a level that breaks your economics, or made subject to new terms you cannot accept. Each of these is a different mechanism, but from your side they all produce the same result: a tool you relied on is no longer usable, and your automation stops.
This is the same lesson, reached from a different direction, that we drew from the largest merger in AI history in our piece on what the SpaceX-xAI deal means for small business. Whether the disruption comes from a merger, a government, a price change, or a product retirement, the defensive posture is identical. You cannot control the events, so you make your business resilient to them instead, and that resilience is a matter of how you build, not how well you guess.
Why this is not a rare one-off
It would be comforting to treat the June episode as a freak event unlikely to recur, but the honest reading is the opposite. AI is now squarely a matter of national interest, with governments around the world increasingly treating frontier models as strategic assets subject to export controls, security reviews, and regulation. That environment makes sudden availability changes more likely over time, not less, because the number of parties who can affect whether a model is usable, and the number of reasons they might act, are both growing.
Beyond government action, the ordinary commercial life of AI models produces regular availability changes on its own. Models are retired and replaced at a rapid clip, and a version you built around this year may be deprecated next year, forcing a migration whether you planned one or not. Prices change, sometimes sharply. Access terms evolve. New regional restrictions appear as laws like the ones we covered in our EU AI Act guide reshape what can be offered where. None of these are disasters if you are prepared, but all of them are disruptions if you are not.
The practical conclusion is that availability change is a normal feature of the AI landscape rather than an exceptional shock, and a sensible business plans for it as a matter of routine, the way it plans for a supplier occasionally raising prices or discontinuing a product. The June event was unusually dramatic because a government was involved and the timing was so tight, but the milder versions of the same thing, a model retired, a price raised, a region restricted, happen constantly and quietly, and they are the ones most likely to catch an unprepared business off guard.
Who is actually exposed?
Not every business needs to worry about this equally, and it is worth being clear-eyed about where you actually stand rather than absorbing a generic anxiety. If your use of AI is light and casual, your team occasionally typing into a chat tool for help with drafting or research, your exposure is minimal, because if one tool became unavailable you would simply open another and lose nothing but a few minutes of adjustment. Most of the disruption risk lands on businesses that have built real operational dependence, where a specific model powers something that runs continuously and matters.
The genuinely exposed businesses are those where an AI model sits in a critical path: a support system that answers customers around the clock, an automation that processes orders or quotes, a workflow that qualifies leads before they reach a salesperson. In those cases the model is not a convenience a person reaches for, it is load-bearing infrastructure, and its sudden absence does not cost a few minutes, it takes down a function the business runs on. If that describes any part of your operation, availability risk is real for you and worth deliberately managing.
Notice that being exposed is not a mistake or a sign you did something wrong. Building real automation on AI is exactly what a forward-looking business should do, and the operational dependence that creates is the natural price of the benefit. The error is not depending on AI, it is depending on one specific model in a way that has no fallback, so that a disruption you could have absorbed instead becomes a disruption that stops you. The fix keeps the benefit while removing the fragility, which is the best kind of fix.
How to protect your business
The core protection is a design principle we return to constantly because it solves so many problems at once: keep the logic of what you are automating separate from the specific model that executes it. When your automation is built so that the model is a swappable component rather than a hardwired foundation, moving from one provider to another becomes a small, quick change rather than a rebuild. That single property turns a model outage from a crisis into an inconvenience, because you simply point the same workflow at a different, comparable model and carry on.
The second protection is knowing your fallback in advance rather than scrambling for one under pressure. For each critical AI-powered function, you should be able to name the alternative model you would switch to if your primary became unavailable, and ideally you should have confirmed that the switch actually works by testing it at least once. The rise of strong, cheap alternatives, including capable open-weight models like the one we covered in our GLM-5.2 explainer, means a viable fallback almost always exists today, so the only real question is whether you have identified and tested yours before you need it.
The third protection is proportion. You do not need elaborate redundancy for a trivial automation, and over-engineering resilience for something that barely matters is its own kind of waste. Match the protection to the stakes: for the AI that customers touch or that your revenue flows through, build in portability and a tested fallback, and for everything minor, accept that a rare disruption is survivable and move on. This is exactly the kind of exposure mapping our €49 audit is built to do, separating the automations that need real resilience from the ones that do not, so your effort goes where it actually protects you.
The bottom line
Two excellent AI models vanished overnight by government order in June 2026 and came back a few weeks later, and the whole episode is worth far more to you as a lesson than it cost anyone as a disruption. It made visible a risk that is usually invisible: the AI you depend on can become unavailable for reasons entirely outside your control, and as governments treat frontier models as strategic assets and vendors retire and reprice models routinely, those disruptions will keep happening in forms both dramatic and mundane.
The response is not to trust AI less or to avoid building real automation on it, both of which would cost you far more than the risk ever will. The response is to build so that no single model's disappearance can break you: separate your automation's logic from the model that runs it, know and test the fallback for anything critical, and size your resilience to the stakes. Do that, and the next time a model goes dark, whether by government order, price hike, or quiet retirement, your business shrugs and switches engines while your competitors scramble. That is the entire payoff of taking a two-week outage seriously as a lesson.
Sources
- Anthropic — Statement on the US government directive to suspend access to Fable 5 and Mythos 5
- Forbes — Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order
- Fortune — Anthropic disables Fable and Mythos AI models following U.S. government export ban
- Al Jazeera — US orders Anthropic to disable AI models for all foreign nationals
- Greenberg Traurig — Anthropic Suspends Access to Claude Fable 5, Mythos 5 Following US Export Control Directive
- CNBC — Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5
- Anthropic on X — Statement on the export control directive