Anthropic Details Disruption of Claude Misuse and Model Distillation

Anthropic has published a threat intelligence report detailing its actions to disrupt the misuse of its Claude AI model. [1] The September 2026 report outlines how the company stopped efforts to use Claude for activities including cyberattacks, surveillance, and influence operations. [1] Additionally, the document details model distillation efforts conducted by Chinese companies. [1] Specifically, Anthropic noted that organizations like Moonshot and DeepSeek were sending user queries to Claude. [1] To accomplish this, these companies routed the queries through "transfer stations" located outside of China. [1]

The threat intelligence report highlights the ongoing challenge AI developers face in securing their models against both state-linked misuse and unauthorized distillation by competitors. The use of proxy "transfer stations" by companies like Moonshot and DeepSeek illustrates the lengths to which international actors will go to leverage restricted frontier models like Claude for training their own systems. What the sources don't address: How many total accounts Anthropic suspended in connection with these operations, or what specific technical countermeasures it has deployed to prevent future proxy access.