Anthropic forced into global model shutdown after U.S. export order targets jailbreak risk

What happened

On June 10, a well-known figure published a jailbreak method on X that appeared to bypass the safety guardrails of Anthropic’s Claude Fable 5, which had launched just days earlier as the company’s most powerful public model. The post claimed to have defeated the model’s safety controls, drawing immediate attention from safety researchers and, according to Anthropic, U.S. officials.

Two days later, on Friday, June 12, the Trump administration abruptly ordered Anthropic to block all foreign nationals from accessing Fable 5 and its sibling model Mythos 5. The directive’s scope was unusually broad: it covered foreign nationals both outside and inside the United States, including Anthropic’s own employees. Because enforcing a nationality-based restriction across a shared, multi-tenant cloud service is technically impractical—API keys do not typically carry citizenship metadata—Anthropic chose the path of least resistance and disabled both models worldwide. The result was an immediate, total blackout of the company’s newest systems for every user, not just those targeted by the order.

Anthropic reviewed the jailbreak report it believes prompted the action and concluded that the capability demonstrated is already widely available from competing models, specifically citing OpenAI’s GPT-5.5. The company has called the government’s response disproportionate and is now working to restore access while attempting to comply with the export controls.

Why it matters

This appears to be the first U.S. export-control measure aimed directly at specific AI models rather than at the chips, hardware, or training infrastructure that powers them. That shift carries significant implications for how AI labs operate. If model weights and inference endpoints themselves can be classified as export-controlled goods based on behavioral risks like jailbreak susceptibility, labs face a new category of regulatory uncertainty that sits upstream of traditional semiconductor controls.

The incident also exposes a structural fragility in the current AI stack. Most leading labs deliver their models through shared cloud APIs where users are authenticated by account and payment method, not by passport or visa status. A sudden order to filter by nationality forces a choice between building rapid, error-prone identity-verification infrastructure or pulling the plug entirely. Anthropic chose the latter, suggesting that compliance architectures for frontier models do not yet exist at the speed regulators are now moving.

Finally, the episode raises questions about proportionality and consistency. If the jailbreak capability Anthropic was cited for is, as the company claims, already present in rival systems that remain online, the action risks appearing both selective and ineffective while doing little to remove the underlying capability from circulation.

Public reaction

No strong public signal was available in captured discussion channels at the time of publication.

What to watch

Observers should monitor whether Anthropic restores access through a nationality-verification layer or negotiates a narrower interpretation of the order. The timeline for restoration remains unclear, and any technical solution will likely set a precedent for how other labs handle similar directives.

It is also worth watching whether other frontier labs—particularly those with models comparable to Fable 5—receive similar orders, which would signal a broader policy shift toward model-level export controls rather than the hardware-focused restrictions seen to date. Finally, any legal challenge from Anthropic or industry groups over the proportionality of the order could define the boundaries of executive power over software exports for years to come.

Sources

• The Verge – "Anthropic got hit by export rules nobody understands"
• The Next Web – "US order to block foreign access to Anthropic’s top models marks a reversal"