Anthropic Denies Chinese Think Tank Access to Mythos AI, Citing US Security Concerns and Tech Rivalry

Anthropic's decision to deny a Chinese think tank access to its powerful new AI model, Mythos, is a direct consequence of escalating U.S.-China tech tensions and growing national security concerns. This move shouldn't be seen in isolation; it's deeply intertwined with a broader shift in U.S. policy. The White House has been actively considering an executive order for pre-release government reviews of frontier models, and Anthropic's decision aligns perfectly with this 'security-first' approach. Granting special access to a Chinese-affiliated entity would have directly contradicted the direction of U.S. policy. So, what specifically pushed them to this conclusion? First, the inherent risks of the Mythos model itself played a significant role. The model was kept under tight wraps following evaluations by both the U.K.'s AI Safety Institute and Anthropic's internal red-teaming efforts, which highlighted novel cybersecurity capabilities. The situation was made worse by reports that a small group of unauthorized users had already briefly accessed Mythos through a vendor. This incident turned the risk of model leakage from a theoretical possibility into a tangible threat, making any new access grant a major gamble. Second, there's a history of mistrust based on past actions. Anthropic had previously exposed what it called 'industrial-scale' distillation campaigns by several China-based labs. These labs allegedly used thousands of fraudulent accounts to interact with Anthropic's other models in an attempt to copy their capabilities. This backdrop changes the interpretation of a 'research request' from a benign inquiry into a potential vector for technology theft. Ultimately, the combination of these factors created a situation where rejection was the only logical outcome. The recent push for government vetting, concrete evidence of leaks, and documented attempts at capability extraction all culminated in this decision, which is a continuation of a policy Anthropic established as far back as September 2025. - Frontier Model: A term for the most advanced and powerful AI models at the cutting edge of technology. - Distillation (AI): A technique where a compact 'student' AI model is trained to replicate the performance of a larger, more complex 'teacher' model. - Red-teaming: A security testing method where a team acts as an adversary to find and exploit vulnerabilities in a system before real attackers do.