When the frontier gets recalled

I build on these models. Vision AI runs on Claude. This site was scaffolded with Claude Code. So when a frontier model gets switched off with no warning, I’m exactly the kind of person who feels it downstream — and yesterday, one did.

Anthropic published a statement that the US government, citing national security authorities, issued an export-control directive to suspend access to Fable 5 and Mythos 5 — its newest and most capable tier. The order targets access by any foreign national, but the practical effect was blunt: to comply, Anthropic had to disable both models for everyone. The rest of the lineup — Opus, Sonnet, Haiku — keeps running.

Independent reporting attributes the order to the Commerce Department’s export-control authority — the same apparatus that polices advanced-chip exports — covering use outside the US and by foreign nationals inside it. That squares with Anthropic’s account.

what they say happened

Per Anthropic’s own account: the directive landed at 5:21pm ET with no specifics attached, and their understanding is that the government believes someone found a way to “jailbreak” Fable 5. Anthropic says it reviewed a demonstration and what it surfaced were “a small number of previously known, minor vulnerabilities” — simple enough that other public models turn them up too, no bypass required.

The specific technique, by their description, amounts to asking the model to read a codebase and fix the software flaws in it. Which is — and I say this as someone who spends his days thinking about what these tools are for — a thing that defenders do every single day.

the part that actually interests me

Lately, a lot of what I build is about making AI admit what it doesn’t know. The CITE method I built CasePrompts on is one long argument that AI should be held to a standard and made to show its work. So I’m not in the camp that thinks safety review is theater. I think the government should be able to stop an unsafe deployment.

But Anthropic’s framing of how that ought to work is hard to argue with:

…a statutory process that is transparent, fair, clear, and grounded in technical facts.

That’s the same bar I’d want applied to any record I produce. A narrow, contested finding that pulls a model used by hundreds of millions of people — before the technical case is even shared — doesn’t clear it. Anthropic says it’s complying while it disagrees, that it believes this is a misunderstanding, and that it’s working to restore access. More is supposedly coming within the day.

So this is a snapshot, not a verdict. By the time you read it the facts may have moved. But it’s worth marking the moment: the tools a lot of us quietly build on are now close enough to the national-security conversation that they can go dark between one afternoon and the next. Worth remembering the next time you wire your whole stack to a single model.

— Andrew