The U.S. government has ordered Anthropic to shut down access to its two most advanced AI systems, Claude Fable 5 and Claude Mythos 5, in a sweeping move that has stunned the AI industry and ignited a fierce debate over how far national security concerns should reach into commercial technology.

Anthropic says it received the directive late in the day and complied immediately, disabling both models for users worldwide. The order was formally framed as an export control measure aimed at limiting access by foreign nationals, but in practice it has yanked the models from every customer, including major U.S. tech firms and security companies.

The stakes are unusually high because Mythos 5 is not a typical chatbot. Internally, Anthropic has treated it as a kind of digital zero-day hunter, capable of uncovering software vulnerabilities at a scale and speed that even elite security teams cannot match. Rather than release it broadly, the company confined Mythos to a tightly controlled initiative known as Project Glasswing, giving vetted partners such as Amazon, Apple, Google, Microsoft, and CrowdStrike access for defensive cybersecurity work.

Fable 5 was Anthropic’s attempt to commercialize that power safely. Built on the same underlying technology as Mythos, it added aggressive guardrails to block high-risk outputs in areas like cyber offense and advanced biology. Independent benchmarks quickly crowned it the most capable publicly available AI model, and Anthropic touted it as proof that cutting-edge capability and rigorous safety could coexist.

According to Anthropic, U.S. officials now believe Fable 5 can be “jailbroken” under narrow conditions to perform vulnerability discovery beyond what regulators are comfortable with. The company says it has been shown only verbal descriptions of a “potential narrow, non-universal jailbreak” involving targeted prompts on a specific codebase, and argues that comparable capabilities already exist in rival systems and open-source tools widely used by security professionals.

Anthropic’s leadership is pushing back hard, warning that if a single edge-case jailbreak is enough to trigger a recall, no frontier model from any major provider would survive similar scrutiny. The company also stresses that its most sensitive safeguards run through independent classifier systems that sit outside the core model, designed to block the very kinds of catastrophic misuse regulators fear.

The shutdown lands at a delicate moment for Anthropic, which has built its brand around being the “safety-first” alternative in the AI arms race and is widely seen as a candidate for a near-term public offering. Now, the same rhetoric that helped distinguish it from rivals is being cited as justification for unprecedented government intervention. In warning the world that Mythos was too dangerous for general release, Anthropic may have convinced Washington that its own creations cannot yet be trusted in anyone’s hands, including its own.