US Blocks Access to Anthropic’s Most Powerful AI Models Over Security Concerns

US authorities have ordered Anthropic to immediately restrict access to two of its most powerful AI models—Claude Mythos 5 and Claude Fable 5. The directive, issued Friday evening, applies to all users worldwide, not just foreign entities as initially expected under export controls.

Mythos 5, released in April, was Anthropic’s most advanced model, capable of detecting vulnerabilities in core operating systems and browsers. Due to high risks, the company limited its use under Project Glasswing, granting access to only about 50 vetted organizations, including Amazon, Apple, Google, Microsoft, and CrowdStrike. Fable 5, launched three days ago, was adapted for broader use with additional safeguards but maintained high performance.

Anthropic stated that the government’s decision stemmed from concerns over Fable 5’s susceptibility to jailbreaking, though the company called this an overreaction. It noted that similar capabilities are already available in other public models, such as OpenAI’s GPT-5.5. Anthropic also emphasized that its safety mechanisms operate independently of the core model, preventing dangerous scenarios even if restrictions are bypassed.

The decision has drawn criticism from Anthropic, which argues that such measures could hinder the entire AI industry’s progress. The company is preparing for an IPO this year and has positioned itself as a safer alternative to competitors. However, its caution around Mythos 5—previously deemed too risky for public release—has now attracted heightened government scrutiny.