Your AI Can Go Offline Overnight. Now What?

At our last Marketo user group I showed an ecosystem diagram with Claude sitting in it as one box among many. This week the US government demonstrated why that one box does not behave like the others around it.

On 12 June, Anthropic took its two newest models, Fable 5 and Mythos 5, offline. The trigger was not demand or a technical fault. It was an export-control directive from the US government, and it is the first time a government has reached into a deployed large language model and switched it off.

If you have written about AI on the assumption that the models only ever get faster, cheaper and more available, this is the week that assumption broke. So let us be precise about what happened, then talk about what it actually means for the way you build.

What actually happened to Fable 5 and Mythos 5?

On 12 June 2026 the US government issued an export-control directive barring every foreign national from Anthropic's Fable 5 and Mythos 5 models, inside or outside the United States, including Anthropic's own foreign-national staff. Because Anthropic cannot verify nationality in real time, it disabled both models for all users. Every other Claude model, including Opus 4.8 and the Claude apps, stays online.

Mythos 5 is Anthropic's specialised cybersecurity model, built to find software vulnerabilities and write the patches for them. Fable 5 is the broader model built on top of it. The government cited a method of "jailbreaking" Fable 5 to reach those offensive capabilities. Anthropic disagrees publicly: it says the technique is narrow, surfaces only minor and already-known flaws, and that the same capability is available from other public models, including OpenAI's GPT-5.5. It is complying with the order while contesting it.

Strip away the cybersecurity drama and the structural fact is the one that matters: a model that hundreds of millions of people could use on Thursday was gone on Friday, by an instruction nobody in the market could see coming or appeal in time.

Is my Claude access affected?

For most people, no. The directive names only Fable 5 and Mythos 5. Claude Opus, Sonnet and Haiku, the Claude apps and the API all keep working. You are affected only if you built specifically on Fable 5 or Mythos 5, or relied on them as a user outside the US, in which case access vanished with no notice and no migration window.

That narrowness is exactly why this is worth your attention rather than your panic. The blast radius this time is small. The precedent is not. A capability can now leave your stack because of where your users hold their passports, and there is nothing in your architecture, your contract or your roadmap that votes on it.

Why this is a Value Gravity™ problem, not just an AI-news story

This is the moment my framework was built to read.

The model is the top layer. It is the brightest, fastest-moving, most exciting part of the stack, and it is also the part you control least. Friday proved the point in the bluntest way available: the top layer can be removed overnight, and not by a market force you could hedge, but by geopolitics and export law. In the three-layer model, durable value does not live up there. It accretes downward, into the layers you own and govern. The model is a swappable, exogenous input that happens to be dressed up as the main event.

There is a sharp edge here for anyone reading this from the Benelux, or anywhere outside the United States. This was a foreign-national export restriction. In that sentence, you are the foreign national. Counterparty risk and sovereignty risk in the AI layer stopped being a slide in a governance deck this week and became a Friday-night email. If your value depends on a model you can be barred from by another country's policy, that dependency is now a board-level fact, not an IT preference.

Offline is offline: the cause barely matters

A model can leave your stack for many reasons: an outage, a price rise, a deprecated version, an acquisition, a licence change, and now a government order. The trigger keeps changing. The exposure does not. The only question that matters is how much of your value sits in a layer you do not control.

It is tempting to file this under "geopolitics" and move on, because an export directive feels exotic and rare. That is the wrong lesson. Anthropic models have had outages this year. Providers deprecate model versions on their own timetable. Prices move. Vendors get acquired. Each of these is the same event wearing a different coat: a capability you depend on becomes unavailable on someone else's schedule. The export ban is simply the most dramatic costume the risk has worn so far, which makes it a free fire drill for a question every enterprise should already have answered.

How do you stop a vanished tool from taking your value with it?

You cannot stop a vendor or a government from pulling a model. You can make sure that when one disappears, your value does not leave with it. That means keeping the things that compound, your data, your logic and your governance, in layers you own, and treating the model itself as a part you can swap.

Six moves, in rough order of leverage:

1. Map your gravity. List the capabilities that depend on a single model at the top layer, and separate them from the ones that live in layers you own: your data, your identity graph, your orchestration logic, your governance. Your exposure is the concentration at the top, and most teams have never drawn it.
2. Keep value in the layer you own. Your data, your prompts, your evaluation sets and your orchestration logic should be portable, not trapped inside one vendor's proprietary surface. If switching models means rebuilding all of that, you do not have a tool, you have a landlord.
3. Abstract the model. Route business-critical flows through a provider-agnostic interface rather than hard-wiring one model's API into them. The goal is that swapping the model underneath is a configuration change, not a project.
4. Qualify a fallback in advance. Have a second model already tested against your prompts and your evaluation sets before you need it. The Friday-night scramble is the most expensive possible time to discover your prompts do not transfer.
5. Make continuity a foundation decision. Treat vendor risk, export risk, data residency and contractual exit as governance, owned at board level, not as a procurement footnote. For any organisation outside the US, jurisdiction is now part of your model-selection criteria.
6. Match dependence to maturity. Let experiments ride the newest, most exciting model, because that is where they belong. Keep production revenue flows on governed, swappable foundations. This is the same discipline I argued for in treating most AI work as laboratory work: the lab can use the concept car, the factory cannot.

So what should a leader take from a Friday-night shutdown?

The excitement lives at the top of the stack. The resilience lives at the bottom. Those are not the same place, and Friday was a reminder of how far apart they can be.

The teams that treated the model as a swappable input read the news, checked their fallback, and got on with their weekend. The teams that had quietly let their value migrate up into a single vendor's model spent it rearchitecting. When the brightest object in the room can be switched off by someone you will never meet, you want your value sitting somewhere with a lower centre of gravity.

That is the whole of it. Don't anchor what compounds to the layer you control least.