Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek

Anthropic released a report on Thursday detailing persistent distillation attacks by China-based AI companies, which have escalated significantly as global competition intensifies. The company stated that unauthorized labs are developing sophisticated methods to circumvent security defenses and harvest the capabilities of US frontier models, specifically targeting Claude’s agentic, coding, data analysis, and logical reasoning features. Anthropic observed nearly 200 million exchanges linked to five separate distillation campaigns. These attacks focus on extracting the "chain of thought" from a model’s response, which is then used to train smaller models via supervised fine-tuning. Although Anthropic restricts access to internal thinking processes, attackers have successfully used prompt engineering—such as framing queries as translation requests—to trick the model into revealing its internal traces. The report highlights that these campaigns are larger and more aggressive than previously identified activity. The most significant effort observed by Anthropic is attributed to Alibaba, which accounted for 151 million exchanges between May and the time of the report, marking it as the largest wholesale distillation effort the company has ever encountered.