Your AI vendor can be switched off by the government overnight. Here's what happened when it actually was.

On 9 June 2026, Anthropic launched Claude Fable 5, its first publicly available model in the new Mythos tier, sitting above Opus in capability. Three days later, the US government ordered it taken offline worldwide. It stayed dark for nineteen days. If your firm, or any client you advise, depends on a single AI vendor for anything that matters, what happened over those nineteen days is worth reading closely, whichever vendor that happens to be.

What actually happened, in order

At 5:21pm Eastern on 12 June, Anthropic received a letter from Commerce Secretary Howard Lutnick, addressed to CEO Dario Amodei, directing the company to suspend all access to Claude Fable 5 and Mythos 5 for any foreign national, anywhere in the world, including Anthropic's own foreign national employees. The directive cited national security authorities and export control powers. Anthropic has said the letter didn't detail the specific concern; its own understanding was that the government believed someone had found a way to bypass safeguards meant to limit the models' use for certain cybersecurity tasks, such as identifying software vulnerabilities.

Anthropic had no practical way to verify user nationality in real time across every platform the models were sold through, AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, Box and its own direct API among them. So it did the only thing actually possible: it switched both models off for every customer, everywhere, that same evening. Every other Claude model, including Opus 4.8, stayed up throughout. Nobody using anything else noticed a thing.

What followed was public disagreement rather than a quiet resolution. A White House adviser was reported as saying Anthropic had been asked to fix the underlying issue and hadn't. Anthropic disputed how serious the jailbreak actually was and was openly critical of how opaque the government's process had been. Amazon's CEO was separately reported to have flagged the issue to officials, worth noting given Amazon's own competing interests in the AI infrastructure market.

On 30 June, the Commerce Department lifted the controls. Lutnick said the department had worked with Anthropic to review and approve Fable 5, in his words to strengthen America's position in AI. Fable 5 came back across Claude Platform, Claude.ai, Claude Code and Claude Cowork on 1 July, with Anthropic giving affected paid plans a temporary 50% top-up on weekly usage limits through 7 July as a goodwill gesture. Mythos 5 access was restored slightly earlier and more narrowly, from 26 June, for a limited set of US organisations, consistent with the tightly gated access it had before any of this started.

The precise national security concern that triggered the order has still never been fully detailed in public. Anthropic's own account is the fullest official version available, and it's worth reading directly rather than through secondhand summaries: anthropic.com/news/fable-mythos-access.

The part worth sitting with

This wasn't an outage, a bug, or a decision Anthropic made on its own. It was a legally binding directive that arrived by letter with well under a day's practical notice, and it took down every customer's access to two specific models worldwide, regardless of whether that customer, or the underlying concern, had anything to do with each other.

Commentators covering the episode made the same point repeatedly: most enterprise contracts and force majeure clauses simply aren't written with "your AI vendor may be legally compelled to switch off overnight" in mind. It's worth actually opening your own vendor agreements and checking how, or whether, a government-mandated suspension is addressed at all. Most firms will find it isn't.

What this actually means for your stack

Not "panic and multi-vendor everything." Most internal, low-stakes tools don't need this level of contingency planning. But for anything client-facing or revenue-critical, four things are worth doing before you need them:

1. Know which workflows are hard-dependent on one specific model versus one provider generally. Most integrations can fail over to a sibling model from the same vendor without much friction. That's exactly what happened here: everything except two named models kept working the entire time.
2. Build and actually test a fallback path, not just document one, for anything that would genuinely hurt if it went dark for three weeks. A fallback nobody has actually switched to is a fallback you don't really have.
3. Read the relevant contracts. Check how force majeure and service-availability clauses in your AI vendor agreements treat a government-mandated suspension specifically, rather than assuming a general outage clause covers it.
4. Match the response to the actual stakes. A firm running a single internal drafting tool has a very different risk profile from one that's built a client-facing product on a single model. Plan for the one that's actually true of you.

Where things stand now

Both models are back and available as of 1 July, and have remained so since. This piece describes events as reported at the time; if you're reading this later and want the current state, Anthropic's own announcement page is the most reliable place to check, and is worth revisiting rather than relying on this summary if anything changes again.

Sources: Anthropic's official statement on the suspension and restoration (anthropic.com/news/fable-mythos-access); reporting from Forbes, CNBC and Lexology/Greenberg Traurig on the directive and its lifting; and enterprise risk commentary from FifthRow and MarketScale on the operational and contractual implications.