Anthropic accuses DeepSeek, Xiaomi, and Moonshot of misusing AI data

Anthropic disclosed on Thursday a report detailing aggressive distillation attacks targeting its AI models by Chinese-based companies. The escalation in recent months has been driven by competition in the AI space. These unauthorized labs have developed sophisticated methods to bypass Anthropic's defenses and harvest capabilities from US frontier models.

The campaigns specifically targeted Claude's agentic capabilities, tool use, coding, data analysis, and logical reasoning. Anthropic identified nearly 200 million exchanges linked to these attacks, attributed to five separate campaigns. The largest wholesale distillation effort observed by Anthropic came from Alibaba, accounting for 151 million exchanges between May and July 2026.

This peaked at nearly three million exchanges per day, targeting 3,500 accounts sharing a single fixed prompt to extract the chain of thought. Another campaign from Moonshot AI, which develops Kimi, seemed to route requests directly from the Chinese military. These requests asked Claude to analyze closed-circuit surveillance footage to determine if subjects were behaving abnormally.

Over a 10-day period, nearly 300,000 requests were routed to Claude through a network of 5,000 accounts, primarily targeting Anthropic's Opus model. Distillation attacks aim to extract the chain of thought from a model's response to queries, which can then be used to train smaller models for general reasoning through supervised fine-tuning.

Anthropic typically does not reveal its models' internal chain of thought, displaying summarized thinking blocks instead. However, these campaigns found techniques to trick the model into revealing its thinking traces directly.