Anthropic launched Claude Fable 5 on Monday. By Thursday night, US export controls killed it.
That’s not an exaggeration. On June 12, at 5:21 PM Eastern, Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei. The directive: suspend all access to Fable 5 and Mythos 5 for any foreign national, anywhere in the world, including Anthropic’s own employees who aren’t US citizens.
Anthropic can’t filter foreign nationals from US users in real time. So they did the only thing they could do to comply. They shut both models down for everybody.
What happened
The timeline here is brutal. Anthropic launched Fable 5 and Mythos 5 on June 9. Fable 5 was the first time Anthropic had made Mythos-level intelligence available to the public. Mythos 5 was the restricted version, limited to vetted partners in Project Glasswing, the cybersecurity coalition that found over 10,000 zero-day vulnerabilities in its first month.
Three days later, both models were dead.
The trigger, according to reporting from Axios and CNBC, was another company claiming to have “jailbroken” Mythos. That claim spooked the Commerce Department enough to act.
But here’s where it gets weird.
The government provided only verbal evidence of the jailbreak. According to Anthropic’s public statement, the technique was “narrow” and “non-universal.” It basically involved asking the model to read a codebase and fix software flaws. 9to5Mac confirmed the directive subjects both models to export controls requiring a license for any domestic transfer to foreign nationals within the US.
Anthropic reviewed it and found the vulnerabilities were “minor,” “previously known,” and “relatively simple.” The kind of stuff any decent security tool can surface.
Then they dropped the real bomb: OpenAI’s GPT-5.5 can do the same thing, without needing any bypass at all.
Why export controls on AI changes everything
Let’s be clear about what just happened. The US government used export controls, a tool designed for missiles, advanced semiconductors, and military technology, to shut down a software product that hundreds of millions of people were actively using. Export controls on software. Let that sink in.
This has never happened before. Previous export controls targeted chip hardware and model weights. This is the first time export controls have hit a live, deployed commercial AI model.
The scope is staggering. It doesn’t just block foreign users outside the US. It blocks foreign nationals inside the US too. If you’re a non-US citizen sitting in a coffee shop in San Francisco, you’re cut off. That includes Anthropic’s own employees.
So think about what that means for every AI company with international staff. If you have engineers on H-1B visas, contractors in London, or a research team in Toronto, you now have to figure out how to separate your user base by citizenship in real time. However, most companies can’t do that. Which is exactly why Anthropic pulled both models for everyone.
Anthropic’s response
Anthropic is complying with the directive while publicly fighting the rationale. That’s a delicate line to walk.
Their statement calls the situation a “misunderstanding” and says they’re “working to restore access as soon as possible.” But they also make a pointed argument: if a narrow jailbreak justifies recalling a commercial product deployed to hundreds of millions of people, then no new frontier model ever ships again.
That’s not hyperbole. Every AI model has non-universal jailbreaks. It’s the nature of the technology. In other words, if the standard is “any jailbreak = government recall,” then OpenAI, Google, Meta, and every other frontier lab is one demonstration away from losing their products too.
Anthropic also points out that its pre-launch testing included thousands of hours of red-teaming with the US government, UK AI Safety Institute, and multiple third parties. No tester found a universal jailbreak. The safeguards were, in Anthropic’s words, “substantially more effective than those of any previously deployed model.”
The political backdrop
This doesn’t happen in a vacuum. The Trump administration has been butting heads with Anthropic for months.
In February, Defense Secretary Pete Hegseth threatened to cut Anthropic out of the DoD supply chain unless the company dropped its AI safety restrictions. Anthropic refused. Trump then ordered US government agencies to stop using Anthropic models.
But here’s the thing. Anthropic also has a partnership with Palantir to provide Claude for classified military missions. The Pentagon’s own CTO reportedly said Anthropic is “still blacklisted” but that Mythos is “a separate issue.”
So you’ve got a company that’s simultaneously providing AI to classified US military operations and getting its best models killed by a different arm of the same government. The contradictions are wild.
What this means for Claude users
If you use Claude Sonnet, Opus, or Haiku, nothing changes. Those models are unaffected.
If you were using Fable 5, you lost access on Thursday night. No warning, no migration path, just gone.
If you’re outside the US, you were already going to lose access under the directive. But Anthropic’s inability to separate users by citizenship means American users got caught in the blast radius too.
Anthropic says it’s working to restore access. The company promised more details within 24 hours of its initial statement (posted around 12:50 AM ET on June 13). Watch for that.
What comes next
This is the story that’s going to define AI policy for the next year. Not because of what happened to Anthropic specifically, but because of the precedent.
If export controls can shut down a live AI product, every frontier lab is now operating under a new constraint. The question isn’t “can we build it?” anymore. It’s “will the government let us ship it?” Export controls just became the most powerful chokepoint in AI. Not compute. Not talent. Access.
For now, the biggest question is whether Anthropic can convince the Commerce Department that the jailbreak threat doesn’t warrant a full product recall. Because the company’s argument is strong: the vulnerabilities are minor, they’re already known, and competing models can do the same thing without any bypass.
But “strong argument” and “government action” don’t always intersect the way you’d expect.
I’ll update this post when Anthropic drops their follow-up statement. If you rely on Fable 5 for work, Sonnet and Opus are still running and worth switching to in the meantime.
