Anthropic reports increased distillation attacks from Chinese AI companies
Anthropic has reported a significant increase in unauthorized distillation campaigns by several China-based artificial intelligence firms, including Alibaba, Moonshot AI, and DeepSeek. These efforts aim to extract the reasoning processes of advanced U.S.-developed models like Claude through sophisticated techniques that bypass security measures. The report highlights over 200 million exchanges linked to these attacks across five distinct campaigns. The distillation process involves tricking large language models into revealing their internal thought chains, which can then be used to train smaller AI systems. Anthropic observed a particularly aggressive campaign from Alibaba involving nearly three million daily interactions and attributed it as the largest such effort seen by the company. Moonshot AI's activities included requests related to surveillance footage analysis. These attacks raise concerns about intellectual property theft in the rapidly evolving field of artificial intelligence, with potential implications for global competition among tech firms.
Topics: AI model security Alibaba AI activities Anthropic report distillation attacks