The United States ordered the global suspension of Mythos 5 and Fable 5, the most advanced artificial intelligence models developed by Anthropic, considering that they represent a risk to national security. The Commerce Department directive forced the company to block access to these systems for any foreign citizen, including its own non-US employees. Faced with the impossibility of verifying the nationality of all users, Anthropic deactivated both models immediately for its entire customer base. The measure marks the first time the US government has imposed export restrictions of this magnitude on a commercial AI model.
The situation worsened when a researcher known as ‘Pliny the Liberator’ managed to breach Fable 5’s security filters in less than 48 hours after its release. Using a multi-agent decomposition technique, the researcher fragmented prohibited queries into seemingly harmless sub-questions that the model processed without triggering its safeguards. Subsequently, an unfiltered version of Claude Opus 4.8 reassembled the pieces to extract detailed guides on cyberattacks, code vulnerabilities and chemical processes for synthesizing controlled substances. The researcher also published on GitHub the model’s complete system prompt, 120,000 characters long, exposing Anthropic’s internal security rules.
Anthropic described the government’s decision as a misunderstanding and warned that applying this criterion to the entire industry would paralyze the launch of new AI models. The European Union, which had obtained access to Mythos 5 after weeks of negotiations, considers that the incident reinforces the need to move toward greater technological autonomy. The case exposes the growing tension between the race to develop ever more powerful AI and the real capacity to guarantee its safe use, at a time when both Anthropic and OpenAI are preparing for their stock market debuts, jointly valued at more than a trillion dollars.
