Anthropic Rejects Chinese Think Tank’s Request for Claude Mythos

Anthropic denied a Chinese think tank’s request for access to its Claude Mythos model during a face-to-face meeting in Singapore in late April or early May 2026.

Anthropic has described Mythos as ‘digital-weapons-grade’ because of its advanced cyber capabilities. The company limits access to roughly 40 institutions in the United States and the United Kingdom, most in defense and intelligence.

The company restricts access to entities with ties to the Chinese Communist Party, including firms under sanctions. Anthropic has estimated that enforcing those limits has cost several hundred million dollars in potential revenue.

Anthropic reported efforts by actors linked to China to bypass its protections. The company identified more than 16 million queries sent via fraudulent accounts that attempted to probe or access restricted functions. In a separate matter, Anthropic attributed an AI-orchestrated cyber-espionage operation to a state-linked Chinese group that it says exploited Claude to target multiple organizations.

Anthropic has placed Mythos behind strict vetting, contractual limits and active monitoring of traffic and account activity. The company states it applies U.S. export controls and its internal safety rules and does not plan a public release of the full-capacity model.

The denial in Singapore followed the company’s established access controls for requests from entities connected to the Chinese state. Other Western AI companies have limited distribution of their most capable systems in contexts involving potential cyber operations or foreign-state actors.