Anthropic's most powerful AI restricted from export by the United States! Fable 5, Mythos 5 go offline worldwide

U.S. Secretary of Commerce Howard Lutnick issued an official letter to Anthropic late at night on June 12 at 5:21 PM, invoking export control authority to prohibit any foreigner from accessing Fable 5 and Mythos 5—two models that had only been live for three days—forcing Anthropic to shut down these two most powerful models directly for all global customers.
(Background: Anthropic released the Claude Fable 5 and Claude Mythos 5 models, setting the strictest safeguards for cybersecurity)
(Additional context: Anthropic CEO: The government should have the authority to veto high-risk AI; models must undergo mandatory testing before launch; three major propositions oppose Trump’s deregulation-drift approach)

Key points summary

• The U.S. Department of Commerce sent a late-night letter on 6/12 banning foreigners from access, forcing Anthropic to shut down Fable 5 and Mythos 5 for customers worldwide
• The fuse was a third-party claim that “jailbreaking” Mythos 5 had been achieved, but Anthropic rebutted that the vulnerability was known long ago and that even GPT-5.5 could do it
• For the first time, the Department of Commerce used export control authority on the AI model itself—exports and transfers both require case-by-case permits

This move by the U.S. government was both swift and severe. At 5:21 PM on June 12, Anthropic received a notice from the U.S. Department of Commerce ordering an immediate prohibition on any foreigners accessing Fable 5 and Mythos 5, which had only been online for three days.

The letter was signed by Secretary Howard Lutnick, with officials from the Bureau of Industry and Security (BIS) behind it. To comply with the requirements, Anthropic had no choice but to directly unplug these two models for “all” customers, cutting off even its own foreign employees.

From launch to global shutdown, it only took three days.

But this emergency stop had actually been brewing for weeks. According to Axios, citing a government official, even before the two models were officially released on June 9, the Department of Commerce was already extremely concerned about the models’ “highly capable cybersecurity reasoning abilities,” having privately tried multiple times to persuade Anthropic to delay the release—without success.

At the time, Anthropic adopted a compromise. What was publicly released was a commercial version of Fable 5 fitted with extremely strict filters: when the model encounters questions touching on cybersecurity, biology, or chemistry, it automatically hands the conversation over to Claude Opus 4.8, which is less capable. The real “safety-lock-free” bare model Mythos 5 was locked inside the tightly controlled “Project Glasswing,” accessible only to about 50 red-team organizations, critical infrastructure operators, and defense units that passed background checks.

These concerns were not unfounded. Anthropic disclosed that within the first few weeks after Project Glasswing went live, Mythos—using 50 partners—found more than 10,000 high-risk and critical vulnerabilities in important software. During testing, if users issued commands, the model could find and exploit zero-day vulnerabilities across every major operating system and every major browser.

“Jailbreaking” surfaced as the reason

What truly lit the fuse was that, just three days after the models were released, a third-party company claimed it had successfully “jailbroken” Mythos 5 in a controlled state and reported the results. For Lutnick and the Department of Commerce, this was the perfect moment for a forceful intervention.

However, Anthropic’s version is somewhat different. The company said the demonstration it reviewed was actually aimed at the filters in the public version of Fable 5. The method involved having the model read a specific piece of code, find and patch software flaws, and ultimately reveal only “a small number of minor vulnerabilities that were already known”—a fairly straightforward approach that even other public models such as OpenAI’s GPT-5.5 can achieve.

Export controls are being used for the first time to restrict the models themselves

This time, the authority the Department of Commerce used was “export control” authority that had previously been used specifically to regulate chips and weapons. Per the letter’s requirements, going forward if Anthropic wants to export, re-export, or transfer these models within the territory, it must apply for permits case by case; any violations will face civil and monetary penalties. This step—using the export control hammer that was originally meant for physical goods and directly swinging it at a software model itself—is unprecedented.

• The ban covers all foreigners inside and outside the United States, including Anthropic’s foreign employees; access is prohibited for everyone
• Permit requirements apply to export, re-export, and domestic transfer of the models; permits must be applied for case by case
• Violations will face civil and monetary penalties
• Only Fable 5 and Mythos 5 are restricted; other Anthropic models are unaffected

Anthropic does not buy this and issued a statement emphasizing that “we believe this is a misunderstanding, and we are working to restore access as soon as possible.” The company also said plainly that if a narrowly scoped potential jailbreaking incident is enough to recall a commercial model that has already been deployed to hundreds of millions of people, then once such a standard is applied across the industry, it effectively means a full stop for the deployment of all cutting-edge models.

We disagree. Discovering a narrow potential jailbreaking should not be a reason to justify recalling a commercial model already deployed to hundreds of millions of users.

An official said these two models may be locked “for the next few weeks,” pending strengthened government national-security support measures before they can be unlocked. Ironically, the day before the shutdown was announced, Anthropic CEO Dario Amodei had just publicly argued that “the government should have the authority to veto high-risk AI.” Now the government has acted—but not in the way he wanted.

Frequently asked questions

Why did the U.S. government freeze Anthropic’s Fable 5 and Mythos 5?

On June 12, Commerce Secretary Lutnick issued a letter citing national security concerns, using export control authority to prohibit foreigners from accessing. The trigger was a third-party company’s claim that it had successfully jailbroken Mythos 5. To comply, Anthropic had no choice but to shut down these two models directly for all global customers.

What are the differences between Fable 5 and Mythos 5?

Fable 5 is a publicly available commercial version equipped with strict filters: when it encounters questions related to cyber security or biological and chemical topics, it automatically hands them over for processing to Claude Opus 4.8. Mythos 5 is a bare model with the safety lock removed; it is made available only through the “Project Glasswing” to roughly 50 reviewed red-team organizations and defense units.