Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek
techcrunch ·
Anthropic released a report on Thursday alleging ongoing distillation attacks by China-based AI companies, which have grown more intense amid heightened competition. The report says unauthorized labs have developed increasingly sophisticated ways to bypass defenses and extract capabilities from US frontier models, targeting Claude features such as agentic and tool use, coding and data analysis, and logical reasoning. Anthropic had previously discussed distillation in February and named specific labs, while OpenAI has reported similar activity attributed to DeepSeek; the new report describes larger and more aggressive campaigns. Anthropic observed nearly 200 million exchanges tied to distillation attacks across five campaigns; the technique seeks to elicit a model's chain of thought to train smaller models through supervised fine-tuning. Although Anthropic normally shows summarized thinking rather than internal chain of thought, attackers used tricks, including a translation-framed prompt, to get the model to reveal reasoning traces. The biggest campaign was attributed to Alibaba, with 151 million exchanges from May to July 2026 peaking near 3 million per day across 3,500 accounts sharing a fixed extraction prompt, linked to training material for Alibaba's Qwen models. A Moonshot AI campaign, tied to Kimi, appeared to route requests from the Chinese military, including one asking Claude to assess surveillance footage for abnormal behavior; roughly 300,000 requests went through 5,000 accounts in 10 days, mainly targeting Opus.
techcrunch