The Anthropic shutdown is a warning to every company building on frontier models: you need a continuity plan, not just a safety policy.

Two frontier AI models were available to businesses one day. Hours after a government directive arrived, they were not.

According to Anthropic, the U.S. government, citing national security authorities, ordered the company to suspend access to its Fable 5 and Mythos 5 models for any foreign national, including its own foreign-national employees. Because Anthropic could not cleanly separate those users in real time, it disabled both models for every customer, not just foreign nationals.

The Dispute Is Not Your Problem. The Outage Is.

You can argue about whether the government was right. It says the models raised a security concern tied to a method of bypassing their safeguards. Anthropic says the evidence it has seen amounts to a “narrow jailbreak”, and that similar capabilities are available from other publicly available models. That dispute will play out over weeks, and most companies deploying AI have no real way to adjudicate it.

Here’s the part they can act on. The model you built a workflow on can disappear overnight, by government order, for reasons your company may not be able to see, test or appeal.

Treat Models Like Infrastructure, Not Software

This is a risk most teams never planned for. Companies have spent two years writing AI safety policies, usage guidelines and acceptable-use rules. Far fewer have written a continuity plan for the day the model itself becomes unavailable. They have been buying frontier models like ordinary software. That is starting to look less like SaaS and more like regulated strategic infrastructure, closer to a power utility or a clearing bank.

Having helped enough companies put AI into production to see the pattern, I can tell you where the exposure usually hides. It’s rarely the marquee use case. It’s the dozens of quiet dependencies: the customer-service flow, the coding assistant, the document pipeline, the internal agent nobody documented. When the model goes dark, all of them go dark at the same moment, and the team finds out how much it had quietly handed to a single provider.

Build The Plan Before The Outage

The fixes are not hard, but they have to be in place before the outage, not after. Keep a second model qualified and ready, not merely named on a slide. Build prompts and tooling to be portable across providers instead of tuned to one.

Then get specific about exposure. Know where your data sits and how long each vendor retains it. Map which workflows touch foreign-national employees or overseas users, because those are now potential points of legal exposure. Put model-availability and notice terms into your vendor contracts. And decide, in advance, which processes fall back to a human when the model is not there.

You don't need a view on the politics to do any of this. It treats government access risk as what it has become: a normal operating variable, like a cloud region failing or a critical supplier going under.

The fight over these specific Anthropic models will move fast. It may be resolved, reversed or half-forgotten within the month. The lesson underneath it will not move. Frontier AI is no longer just a capability you adopt. It’s a dependency you have to govern, and the list of things that can take it away from you now includes your own government.