Anthropic Report on Distillation Attacks by Chinese AI Firms
A newly published report reveals extensive operations conducted by China-based artificial intelligence companies targeting U.S. frontier model capabilities.
A new report published by Anthropic on Thursday revealed that China-based artificial intelligence companies, including Alibaba, Moonshot AI, and DeepSeek, have been carrying out persistent distillation attacks to capture the capabilities of U.S. frontier models.
Scope and Objectives of Distillation Attacks
Anthropic stated that amid the intensifying competitive environment in recent months, unauthorized labs have developed increasingly sophisticated methods to bypass defenses.
The report emphasized that the identified campaigns targeted the Claude model's most valuable capabilities, such as agentic and tool use, coding, data analysis, and logical reasoning.
Millions of Interactions and Large-Scale Campaigns
Anthropic reported observing approximately 200 million interactions across five distinct campaigns connected to distillation attacks.
These attacks aim to extract the chain of thought from model responses in order to train a smaller model in general reasoning capability.
Largest Wholesale Distillation Operation Originating from Alibaba
A significant portion of the attack attempts came from an Alibaba-linked campaign, described by Anthropic as the largest wholesale distillation effort observed to date.
Between May and July 2026, 151 million interactions attributed to this campaign were observed, with activities reaching three million interactions per day.
Moonshot AI and Military-Oriented Requests
Another campaign conducted by Moonshot AI, the maker of the Kimi model, showed that requests were directed straight from the Chinese military.
Over a ten-day period, approximately 300,000 requests, mostly targeting the Opus model, were submitted to the Claude system.