Commentary: Why the US government shut down Anthropic’s latest Claude AI model

The U.S. government suspended public access to Anthropic’s latest Claude AI models, Fable 5 and Mythos 5, just three days after their release on June 9, 2025. The ban, enforced through an export control directive, restricts use of the models to U.S. nationals only. Mythos 5, Anthropic’s most advanced "frontier" model, was initially withheld in April due to its ability to exploit cybersecurity weaknesses, while Fable 5—designed with safeguards to prevent misuse—was publicly accessible before its shutdown. The conflict between Anthropic and the Trump administration has intensified since early 2025, with disputes over AI regulation, semiconductor policies, and defense applications. The Pentagon had threatened to label Anthropic a 'supply chain risk,' forcing military contractors to cut ties, after the company refused to allow its AI models to be used for domestic surveillance or autonomous weapons. The shutdown follows reports of a jailbreak—methods to bypass Fable 5’s safeguards—that could expose sensitive information for cyberattacks. Anthropic confirmed the U.S. government likely discovered such vulnerabilities, though the company acknowledged that no AI model can be fully immune to jailbreaks. A researcher using the pseudonym 'Pliny the Liberator' had already leaked Fable 5’s system prompt within 48 hours of its release, raising further concerns. Anthropic’s safeguards rely on the model’s ability to interpret user requests, but their effectiveness is limited by the complexity of AI systems. The company cited Amazon’s engineers, both rivals and investors, as potential sources behind the government’s findings. Meanwhile, the broader AI community, known as the Undersphere, actively works to exploit these weaknesses, making secure deployment a persistent challenge. The incident highlights broader tensions over AI governance, with the U.S. government tightening controls amid fears of misuse. Anthropic’s models, even with restrictions, have demonstrated capabilities that outperform expectations, raising questions about the feasibility of safeguarding advanced AI systems.