Anthropic Tells China No Access to Its Most Powerful AI Model

Artificial intelligence company Anthropic has refused a Chinese think tank’s request to access its Mythos model, one of the most capable cybersecurity AI systems ever built. The refusal came during a meeting in Singapore last month, where Chinese officials pushed Anthropic to change its stance on blocking access to Beijing, according to the New York Times. The decision underscores how advanced AI models are becoming tools of geopolitical competition.

What Is Mythos and Why It Matters

Claude Mythos Preview launched on April 7, 2026 as Anthropic’s most advanced model for identifying and exploiting cybersecurity vulnerabilities in software. Rather than releasing Mythos publicly, Anthropic distributed it through a program called Project Glasswing. This initiative gave access to a small group of about 50 United States companies, including Cisco, JPMorgan Chase, and Nvidia. The United States National Security Agency and other government agencies also began using the model to help secure critical infrastructure.

Anthropic classifies China as an adversarial nation, which means Chinese companies and organizations cannot use Mythos or any of Anthropic’s services. This ban extends across greater China, including Hong Kong. A large language model, or LLM, like Mythos works by scanning millions of lines of code to find security weaknesses that human researchers might take months to discover. That ability makes it both a powerful defensive tool and a potential weapon.

A Widening AI Gap Between Washington and Beijing

The refusal highlights a growing divide between the two countries in AI-powered cybersecurity. Beijing-based consultancy Concordia AI has noted that China’s mostly open-source AI models still trail behind closed-source American models when it comes to finding and patching security flaws. However, Chinese AI capabilities are advancing rapidly.

The White House has also pushed back against expanding Mythos access even among allies. When Anthropic proposed increasing its distribution from roughly 50 organizations to 120, officials opposed the plan over concerns about potential misuse and the computing power required to run it safely. Even the European Union has not yet received access, with discussions described as being at a different stage than those with other AI industry partners.

The decision to keep Mythos restricted reflects a broader shift in how companies release their most sensitive AI systems. Advanced cybersecurity models now require distribution controls similar to those used for defense technology, marking a new chapter in the relationship between Silicon Valley and national security.