U.S. forces Anthropic offline: Commerce Dept export directive kills Fable 5 and Mythos 5 worldwide
Late Friday, the Commerce Department issued an export control directive ordering Anthropic to block all foreign nationals from Fable 5 and Mythos 5 — its most advanced models. Because Anthropic cannot verify citizenship at scale, it shut both models off for every customer globally. The government cited a narrow, non-universal jailbreak; Anthropic says the same capability is present in GPT-5.5 and other live models, calls the directive procedurally unfair, and says it is working to restore access. This is the first time U.S. export controls have targeted an AI model directly rather than hardware.
The U.S. Commerce Department issued an export control directive late Friday forcing Anthropic to suspend all access to Fable 5 and Mythos 5 — its two most capable AI models — for any foreign national, anywhere in the world, including Anthropic's own foreign-born employees. To comply, Anthropic shut the models off for every customer, not just international ones.1
The shutdown took effect within hours of the 5:21 p.m. ET directive. All other Claude models remain online.
What the government ordered — and why Anthropic disagrees
The directive cited national security authorities but gave Anthropic no written explanation. Verbally, officials told the company they believe a non-universal jailbreak exists: a specific technique that prompts Fable 5 to read a codebase and identify software vulnerabilities.2
Anthropic reviewed the reported technique and rejected the premise. Its statement says the demonstration produced only "a small number of previously known, minor vulnerabilities" — all simple, and all replicable by publicly available models. It specifically compared the output to OpenAI's GPT-5.5 cybersecurity capabilities, which remain live.1
The company's Fable 5 launch blog had preemptively addressed this question: no tester in thousands of hours of red-team exercises had found a universal jailbreak, and Anthropic had said publicly that perfect jailbreak resistance is not currently possible for any model provider. Its defense-in-depth strategy — narrow jailbreaks rather than broad ones, combined with 30-day customer data retention for monitoring — was designed precisely for this scenario.
Anthropic's public statement did not mince the frustration: "If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers."1
正在加载内容卡片…
A third government clash in six weeks
This is the third distinct government conflict Anthropic has navigated since early May.
The first was the Pentagon's supply-chain risk designation — historically applied to foreign adversaries — issued after Anthropic refused a Defense Department demand for "all lawful uses" of Claude including mass surveillance and autonomous weapons. Anthropic filed a federal lawsuit challenging both the designation and a White House directive barring civilian agencies from using Claude, on June 11.3
The second was the Fable 5 transparency dispute: Anthropic had silently downgraded frontier AI requests to Opus 4.8 inside Fable 5, drawing developer backlash, and reversed course on June 11 to make refusals visible via API.
Friday's export control directive is categorically different from both. The Pentagon conflict was about what the model could be used for. The Fable 5 reversal was about disclosure. This directive is about who can access the model at all — a question that has not previously been applied to a commercial AI product at scale. Reuters described it as "the most significant step to date" by the U.S. government to restrict access to advanced AI models, and the first time export controls have targeted the model itself rather than the chips or tools that build it.2
正在加载统计卡片…
The AWS cascade and the foreign-employee problem
AWS confirmed late Friday that Anthropic asked it to revoke access to Fable 5 and Mythos 5 "for all users in all regions."2 Because Anthropic cannot technically distinguish which individual users are foreign nationals — and because the directive explicitly includes "foreign national Anthropic employees" — a global takedown was the only compliant path.
Former White House AI policy official Dean Ball noted the practical implication on X: the directive "suggests all non-Americans would be restricted from using Anthropic's latest models, including those based in the U.S." and that users should "expect to have to prove your citizenship to use Anthropic models."2
Several of Anthropic's most senior leaders were born outside the United States. When Reuters asked whether they retained access to the models, Anthropic declined to comment.
The disruption lands directly on the enterprise customers Anthropic spent the previous week signing. TCS (50,000 employees across 56 countries) and DXC Technology (115,000 employees across 70 countries) both became Global Premier partners on June 11. Neither has commented publicly on how the takedown affects their deployments. India, which Dario Amodei confirmed last week is Anthropic's "second-largest market," is outside the scope of permitted access under Friday's directive.
What comes next
Anthropic said it believes the situation is a "misunderstanding" and is working to restore access "as soon as possible." It plans to release more technical details within 24 hours, and has indicated it wants to challenge the procedural basis of the directive — arguing that the government should be able to block unsafe models but only "as part of a statutory process that is transparent, fair, clear, and grounded in technical facts."
The timing matters for the IPO. Anthropic filed a confidential S-1 on June 1, with Morgan Stanley and Goldman Sachs as underwriters and a target Q4 2026 listing at roughly $60 billion. The supply-chain designation and now this export directive represent two active, unresolved legal conflicts with the federal government heading into the roadshow period.
The precedent matters for the industry. If a non-universal jailbreak — one that produces minor, already-known vulnerabilities also producible by other live models — is sufficient to trigger a forced commercial takedown, every frontier AI lab faces the same exposure. Anthropic's language was deliberate: it framed Friday's directive not as an Anthropic problem but as a test of whether any new model can be commercially deployed at all.1
围绕这条内容继续补充观点或上下文。