On June 9, Anthropic released Claude Fable 5, the most powerful model it had ever made generally available, and Mythos 5, a restricted but equally powerful sibling. They're the same underlying model — same weights, same price — and the only thing separating them is the safety stack that decides what each one will tell you. On June 12 at 5:21pm ET, the US government handed Anthropic an export-control directive. It barred access to both models by "any foreign national, whether inside or outside the United States, including foreign national Anthropic employees." Anthropic can't filter foreign nationals from US users in real time, so it disabled both models worldwide within hours; every other Claude model, including Opus 4.8, stayed up. The reason the government gave: someone had found a way to jailbreak Fable — to get around the safety limits that decide what it will and won't do.

1. This Is Exactly What Export Controls Are For (David Sacks, China hawks)

Mythos is the dangerous one. A jailbroken Fable hands it to everybody.

Fable and Mythos are the same model. Same weights, same capabilities, and the only thing between them is a safety stack that governs what the model will say. Mythos, the version with fewer restrictions, goes only to vetted cybersecurity and biology partners. A jailbreak that strips Fable's safety stack hands the public Mythos.

The trigger wasn't theoretical. Semafor reported the White House acted partly over suspicion that a China-linked group had gotten into Mythos. David Sacks, Trump's AI adviser, said the government was warned Fable could be jailbroken and that Amodei called the risk minor and wouldn't fix it. Amazon CEO Andy Jassy — whose company is a major Anthropic backer — reportedly warned the government about the jailbreak too.

Frontier cyber capability is the whole reason export control exists. Lawmakers are advancing a bill to make curbing AI-model exports easier, and US fear of China using AI in autonomous weapons keeps rising. Security researchers warned at launch that a model this good at writing code is just as good at finding and exploiting the flaws in it.

2. This Is Wildly Disproportionate (Anthropic, Dario Amodei)

The "jailbreak" is a code review. The capability is everywhere already.

The "jailbreak" is just a code review. The scary technique the government pointed to is this: hand the model some code and ask it to find the security holes in it. Programmers and security teams do that every day. When Anthropic looked at the demo, the model only turned up a few small bugs that were already known — and any model can do this, OpenAI's GPT-5.5 included.

Pulling a model over this would freeze the whole industry. Anthropic says yanking a model "deployed to hundreds of millions of people" over "a narrow potential jailbreak" would set a standard that "would essentially halt all new model deployments for all frontier model providers." Anthropic agrees the government should be able to block a genuinely dangerous model. Its problem is how this one happened — overnight, with no hearing and no standard it could answer, nothing "transparent, fair, clear, and grounded in technical facts."

3. Even If You're a Hawk, This Helps China (Dean Ball, Gary Marcus)

It helps China more than it hurts them — the opposite of what it's supposed to do.

The chip policy and the model policy point in opposite directions. Dean Ball, a former Trump AI adviser, called the move "cartoonish" and "baffling": the same administration loosening export controls on advanced AI chips to China just locked every foreign national on Earth out of a chatbot. He asked whether it was lawfare or "extreme national-security hawkery."

Lock foreign researchers out and they go home — to China. Gary Marcus argued the action made little sense for US competitiveness and would push Chinese-born researchers back to China, which is the exact opposite of what a hawk should want.

4. The Problem Is There's No Playbook (Justin Hendrix, Brennan Center, R Street)

Maybe the concern is real. They made the method up on the spot.

The government recalled a deployed model with no process behind it. Justin Hendrix of Tech Policy Press argued the US shut down a live model case-by-case, with no transparent framework, which leaves allies unable to trust American AI and invites "picking favorites based on personal and political factors." The Brennan Center's Amos Toh called the directive "a study in conflicting impulses."

Even the sympathizers didn't like the execution. R Street's Adam Thierer said the concerns were legitimate but the action was "a significant escalation in the politicization of AI and centralization of control over advanced computation."

5. Your AI Can Vanish Overnight (builders, affected subscribers)

The most capable model on Earth disappeared in hours. Users had no say.

This is the cleanest argument for local AI anyone's made all year. The most capable model Anthropic had ever shipped went dark in hours — not over price or performance, but a federal order nobody outside the room could see coming. Anyone who built a workflow on Fable on Wednesday had no model on Friday.

The refund process shows how little control users actually have. Anthropic opened a cancellation window into late June, but you have to do it from a desktop browser, Apple subscribers have to go through Apple, and the refunds have come back partial and contested. The model you paid for is gone, and getting your money back is a separate hassle.

Where This Lands

A federal order pulled the most powerful model Anthropic had ever shipped three days after it launched, and the fight is over whether that was proportionate. The government's defenders say Fable and Mythos are the same model, so a jailbreak hands a cyber-capable system to anyone — including a China-linked group that may have already gotten in. Anthropic says the "jailbreak" is a code review every frontier model can do, and that recalling a live model over it would freeze the industry. Critics on both the hawk side and the rule-of-law side say the goal might be defensible but the method was incoherent and made up as it went. And for everyone who built on Fable, the lesson is simpler: the AI you depend on can be switched off overnight by people you'll never meet.

Sources