Anthropic Got Shut Down Three Days After Launch. Here Is What That Means.

Three days after Anthropic launched Fable 5 and Mythos 5, the Trump administration sent Commerce Secretary Howard Lutnick’s letter to Dario Amodei and the models went dark for every customer, everywhere. Not just foreign governments. Everyone.

The stated reason: another company claimed it could jailbreak Mythos 5, the restricted cybersecurity-focused model available only to vetted US government users via a program called Project Glasswing. That claim was enough for the Commerce Department to impose export controls and demand that Anthropic prevent any foreign national from accessing both models. Since Anthropic has no technical way to verify the nationality of every user in real time, “comply with the directive” meant “shut it down for everyone.”

That is the situation. Here is what I think it actually means.

This Has Happened Before

In the 1990s, the US government classified strong encryption as a munition and tried to control its export under the same legal framework used for weapons. Phil Zimmermann was investigated for making PGP freely available online. Researchers were threatened for publishing cryptography papers. The NSA pushed the Clipper Chip, a hardware backdoor baked into telephones.

The government lost that fight. Not because it gave up, but because math does not respect borders. Once the algorithms were published, they were everywhere. Export controls on cryptography collapsed not from a single court ruling but from the sheer impossibility of enforcing them in a world of global internet connectivity.

The parallel to AI is imperfect but the shape is familiar. The thing the government is trying to control is information, not hardware. And last week alone, twenty-five open-weight models dropped across every modality. While Fable 5 and Mythos 5 are locked behind a Commerce license, Liquid AI’s LFM2.5-8B runs on your laptop and Ideogram 4’s weights are downloadable by anyone on the planet. The capability the government fears is available; it is just not labeled Anthropic.

The Safety Trap

There is a specific irony in Anthropic’s situation worth naming. Anthropic is the AI lab that made safety its founding identity. It partnered with the Center for AI Standards and Innovation at Commerce for pre-deployment testing. It built Mythos 5 specifically for US government cyber defenders, under a restricted access program, with safety guardrails designed to prevent harm.

The result: Anthropic is now simultaneously on a Pentagon blacklist deeming it too dangerous for the government’s own use, and subject to export controls deeming it too dangerous for anyone outside the US. The lab that tried hardest to work with the government is the one that got shut down.

This does not mean safety work is pointless. But it should make other labs think carefully about the assumption that regulatory cooperation creates protection. It can just as easily create exposure. If you sit at the table and let the government test your models under a pre-deployment agreement, you also hand them the mechanism to act when they find something they dislike.

What “Jailbreak” Actually Means Here

The word jailbreak is doing a lot of work in this story and it is worth being precise. A jailbreak in this context means someone found a way to get a model to produce outputs it was designed to refuse. That is a real security concern. But the threat model matters enormously.

The government’s argument seems to be that a jailbroken Mythos 5 could provide meaningful uplift to adversaries. That is a testable claim, and Anthropic’s own statement says the demonstration revealed “a small number of previously known, minor vulnerabilities.” A narrow jailbreak on a model that is already restricted to vetted US users is a very different risk class than, say, a bioweapons synthesis route freely available through an API.

The Commerce Department appears to have treated a commercial safety incident the way it would treat a weapons system vulnerability. Those are different things. Conflating them produces policy that does real harm to real users, without meaningfully reducing the actual risk, because the underlying capability is available through other means.

What This Means If You Build on AI APIs

The practical lesson is blunt: the ground can shift overnight. Seventy-two hours after launch, with no warning, every customer lost access to two models simultaneously. If your product depends on a specific model from a single provider, you have a single point of failure that your infrastructure runbooks do not cover.

This is not unique to Anthropic. Any sufficiently advanced closed model is a candidate for export controls under the logic the administration applied here. The more capable the model, the more it looks like a national security asset.

For teams building on AI APIs, the architecture question is not just “which model is best?” anymore. It is “what is our fallback if this model becomes unavailable?” The answer increasingly points toward model-agnostic abstraction layers, multi-provider routing, and at least some capacity to run open-weight models locally when the API goes dark.

The irony is that the government’s action accelerated exactly the behavior it was trying to prevent: driving sophisticated users toward open-weight alternatives that nobody can export-control.

Where This Goes

The administration says the lockdown could lift “in the next few weeks” once the national security apparatus is hardened. Anthropic says it is a misunderstanding and is working to restore access. Both of those statements might be true.

But the precedent is set. The US government has demonstrated that it will treat advanced AI models as export-controlled technology, that a single jailbreak claim is sufficient to trigger action, and that commercial impact on millions of users is not a blocking consideration.

If you think that precedent stays limited to Anthropic, or stays limited to the current administration, I would not bet on it.

---

Source: Axios — Scoop: Trump admin blocks foreign access to Anthropic’s most powerful AI