US Government Order Suddenly Blocks Anthropic's Latest Models Worldwide

Last week, the US government issued an unprecedented export control directive ordering Anthropic to immediately suspend access to its top-tier AI models "Claude Fable 5" and "Claude Mythos 5" for foreign users. While cited on national security grounds, specific details have not been disclosed.

In response to this directive, Anthropic has completely blocked public access to both models globally. No exceptions are made for paid corporate customers or even internal employees, leaving the models inaccessible to everyone worldwide. This occurred just three days after Fable 5 and Mythos 5 were released, dealing a significant blow to the company.

Currently, sessions for both models terminate with errors, and new queries are automatically rerouted to legacy models such as "Opus 4.8." Anthropic stated in an official blog post that "we believe this measure is based on a misunderstanding and are working as quickly as possible to restore access," offering an apology to customers.

Attracting attention regarding the background of this government intervention is the public disclosure of attack methods by renowned jailbreaker "Pliny the Liberator." On June 10, he published detailed techniques on X (formerly Twitter) claiming to have breached Fable 5's safety measures. He stated that he successfully extracted information on cyberattack procedures, explosive device manufacturing methods, and methamphetamine synthesis using the "birch reduction method." The attack reportedly combined Unicode and homoglyph characters, Cyrillic script mixing, long-context reference tracking, and techniques for fragmenting harmful requests into harmless tokens.

However, Anthropic has not explicitly stated whether this attack directly triggered the government order. Rather, it explained that "we only received oral explanations from the government and were only informed about the possibility of a non-universal jailbreak that involves reading specific codebases and fixing bugs." Furthermore, Anthropic pointed out that the problematic functionality is "widely available" in other publicly released models such as OpenAI's "GPT-5.5," and strongly warned that suspending commercial models due to non-universal jailbreaks could "effectively halt new model deployments by all frontier model providers."

This incident has provided important lessons for enterprises. Cloud-based frontier AI models are directly subject to government regulations and vendor compliance measures. The risks of concentrating critical business operations on a single AI provider or model have been highlighted once again. Enterprises now face an urgent need to consider redundancy and diversification across multiple models and alternative AI solutions.