The U.S. told the AI company to lock out every foreign national from its newest models, citing a technique for tricking chatbots into ignoring their safety rules. Anthropic says the technique surfaces only minor flaws.
At 5:21 p.m. ET on Friday, Anthropic received a phone call from the U.S. government. By Saturday morning, every user outside the United States, and every foreign national inside it, had been cut off from Claude Fable 5 and Mythos 5, the company's two newest and most capable models.
The order arrived as an export-control directive, a national-security tool the U.S. uses to keep controlled technology out of foreign hands. A technique for tricking an AI into ignoring its safety rules, a so-called "jailbreak," had allegedly made the models a national-security risk, the company said. Anthropic disabled access the same evening. Other Anthropic products, including earlier Claude models, remain available.
Anthropic disputes the government's reasoning. The company says it reviewed a government-supplied demonstration of the Fable 5 jailbreak and found only a small number of previously known, minor vulnerabilities. Other publicly available AI models, the company added, are sufficient for the same tasks. "We believe this is a misunderstanding," Anthropic said, "and we are working to restore access as soon as possible" (The Hacker News, reporting Anthropic's statement).
The U.S. government has not published the order. No agency has gone on the record to confirm it issued the directive or to describe its scope. The public record, for now, is a company statement, a security outlet's summary, and an unexplained decision that, in a few hours, removed frontier AI access from a global user base.
Fable 5 is Anthropic's newest general-purpose large language model. Mythos 5 is its cybersecurity-focused model, shipped with some safety guardrails lifted in defensive-security domains and reserved for vetted defenders and critical-infrastructure operators. Together they represent Anthropic's frontier tier. The order, as the company describes it, treats those models less like commercial software and more like controlled technology that can be denied to non-U.S. users on national-security grounds.
Frontier AI models have not historically been placed under U.S. export controls this aggressively. Past rules have governed chips, model weights above defined compute thresholds, and specific foreign end users. An export-control directive that disables consumer-facing model access for foreign nationals on hours of notice is a different kind of action, and one with a much wider footprint.
What the U.S. is actually worried about remains unclear. If the concern is that the jailbreak can extract functional cyberattack code from Fable 5, Anthropic's response is that the same prompts work on other public models. If the concern is that foreign adversaries could run the model directly, the order reaches only Anthropic's hosted product, not the underlying research. Without the order's text, the scope, and the legal hook it relied on, the public argument is mostly Anthropic's.
The story is moving. Watch for an Anthropic blog post or status-page update, an on-record statement from a U.S. agency such as Commerce's Bureau of Industry and Security or the National Security Council, and any follow-on action against other frontier model providers. The narrow technical question is whether the Fable 5 jailbreak surfaces anything the company's peers cannot also surface. The wider question is whether the U.S. will now treat a frontier chatbot as something it can pull from foreign hands on a Friday afternoon.