On 12 June 2026, a frontier AI model was running for hundreds of millions of people. By the evening, it was gone.
The US government ordered Anthropic, the company behind Claude, to suspend access to two of its models, Fable 5 and Mythos 5. Anthropic received the letter at 5:21pm Eastern time and had to switch both models off for every customer in the world to comply. No long notice. No phased wind-down. Off.
If you run a business that uses AI, or you are thinking about building something on top of it, this is the story to pay attention to. Not because the sky is falling, but because it shows a risk most people never priced in: the model you depend on can be taken away by someone who is not your vendor and not you.
Here is what happened, in plain English, and what it means for you.
What actually happened
Anthropic builds AI models. Two of its most capable ones are called Fable 5 and Mythos 5. We wrote about what these models can do when they launched, in Claude Fable 5 and Mythos 5: what Anthropic's most capable model means for your business.
The US government issued what is called an export control directive. In short, it is a legal order that restricts who can access a piece of technology, often on national security grounds. This one said no foreign national, anywhere in the world, could use Fable 5 or Mythos 5. Because Anthropic could not cleanly separate "foreign nationals" from everyone else fast enough, the only way to obey the order was to turn the models off for all customers.
Two things are worth being clear about. First, this hit only those two models. Anthropic stated plainly that access to all its other models, including the Claude models most businesses actually use day to day, was not affected. Second, Anthropic does not agree with the order. The company said it is complying because it is the law, but it believes the decision is a misunderstanding and is working to get access back.
The reason given, and Anthropic's response
The government pointed to national security. Its specific concern, as Anthropic understands it, was a way of "jailbreaking" Fable 5. A jailbreak is a trick that gets an AI model to do something its safety rules are meant to block.
Anthropic says it looked at the demonstration. Its read: the technique surfaced a handful of minor software flaws that were already known, and that other freely available AI models can find the same flaws without any special trick. The company named OpenAI's GPT-5.5 as one example of a model with comparable ability. In Anthropic's words, this is the kind of thing the people who defend computer systems do every day.
Anthropic also laid out how it had prepared. Before launching Fable, it ran red-team testing for thousands of hours with the US government, the UK's AI Safety Institute, outside organisations, and its own teams. Red-teaming is when you pay skilled people to attack your own product before anyone else does, so you find the holes first. Anthropic said no tester had found a universal jailbreak, meaning a single trick that broadly defeats the model's safety rules. It even kept 30 days of customer data specifically so it could spot and shut down abuse, a policy it admits costs it business with privacy-conscious customers.
The company's bottom line: pulling a model used by hundreds of millions of people over one narrow, unconfirmed jailbreak is the wrong call, and if that became the standard, it would freeze new model releases across the whole industry. Anthropic said it would share more detail within 24 hours, so some of this may look different by the time you read it.
I am not here to settle who is right. The point for your business sits one level up.
The real lesson: you do not control the off switch
Most businesses think about their AI vendor the way they think about their electricity supplier. You pay the bill, the service stays on, and the only real risk is a price rise. This event breaks that assumption.
A model you build on can disappear for reasons that have nothing to do with you, your contract, or even your vendor's wishes. A government can order it. A regulator can restrict it. The vendor can retire it to focus on a newer one. Any of these can happen with little warning.
This matters more, not less, if you are in the GCC or Australia. The order in this case targeted foreign nationals specifically. If your business sits outside the United States, you are exactly the kind of user a future order like this could touch first. You are also further from the decision, with less ability to influence it or even get early word that it is coming.
Call this AI vendor risk: the chance that the model your product runs on becomes unavailable, restricted, or far more expensive, for reasons outside your control. It is not a reason to avoid AI. It is a reason to build so that one shutdown does not take your business with it.
What to do about it, in plain terms
You do not need a crisis plan. You need a few sensible habits, most of which your developer or agency can put in place quietly.
Keep the AI provider behind a thin layer in your own software. In practice this means your app talks to a small piece of your own code, and that code talks to the AI provider, rather than your features calling the provider directly all over the place. When you need to switch models, you change one spot, not fifty. We use this approach on client builds for exactly this reason.
Avoid welding one specific model into the core of your product. The more a single model's exact behaviour is baked into your most important features, the more painful it is to move. Treat the model as a part you can replace, like a supplier you could change, not a foundation you pour concrete around.
Know your backup before you need it. Pick a second model from a different provider that could do the job in a pinch, and confirm your system can actually point at it. You do not have to run it. You just have to know the switch exists and works.
Match the model to the sensitivity of the work. The models pulled here were Anthropic's most capable, the ones aimed at the heaviest tasks. A lot of real business work, drafting, summarising, answering customer questions, runs perfectly well on lighter, broadly available models that are less likely to attract this kind of attention. Use the heavy model where you truly need it, not everywhere by default.
If you only do one thing, ask this question out loud at your next planning meeting: if our main AI model vanished tomorrow morning, what stops working, and how long would it take us to switch? If nobody can answer, that is the gap to close. For a fuller picture of how AI fits a business safely, 13 new AI terms every founder should actually understand in 2026 and AI agents vs chatbots: why 2026 is the year of autonomous AI are good places to start.
What to do now
This story will keep moving, and access to Fable 5 and Mythos 5 may well be back by the time you read this. The headline event is less important than the habit it should leave you with: treat the AI model under your product as something that can change without your say-so, and build so that a switch is a small job, not a rebuild.
If you are not sure how exposed your current setup is, that is worth a short conversation. At dsrpt we build AI into client products across Australia and the GCC with this exact risk in mind, so a model going dark is an inconvenience, not a shutdown. If you want a straight answer on where your business stands, get in touch.