Anthropic exposes distillation campaigns by chinese AI firms

Anthropic has disclosed ongoing distillation attacks by Chinese AI companies, which have intensified in recent months amid heightened competition. The report highlights sophisticated methods used by unauthorized labs to bypass defenses and extract capabilities from US frontier models, particularly targeting Claude’s agentic capabilities, tool use, coding, and logical reasoning. Anthropic previously raised concerns in February, with OpenAI attributing similar activity to DeepSeek. The latest report details larger and more aggressive campaigns, with nearly 200 million exchanges linked to distillation attacks across five campaigns. Distillation attacks focus on extracting a model’s chain of thought to train smaller models through supervised fine-tuning. Anthropic noted that attackers found ways to trick models into revealing thinking traces, such as framing queries as translation requests. Alibaba’s campaign, attributed as the largest distillation effort, involved 151 million exchanges between May and July 2026, peaking at three million daily exchanges. Moonshot AI’s campaign, linked to the Chinese military, routed nearly 300,000 requests to Claude’s Opus model over ten days via 5,000 accounts.