Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek (techcrunch.com)

Anthropic has revealed significant and alarming findings regarding aggressive distillation attacks conducted by several AI companies based in China, including Alibaba, Moonshot AI, and DeepSeek. In a report published Thursday, Anthropic noted that these unauthorized labs have developed advanced methods to extract valuable capabilities from U.S. AI models, particularly their reasoning abilities, coding skills, and tool usage. Over a series of campaigns, Anthropic identified nearly 200 million exchanges of this information, with Amazon's Alibaba leading the charge in what the company describes as the largest distillation effort observed to date. The implications of these attacks are profound for the AI/ML community, as they expose vulnerabilities in how frontier models protect their internal reasoning processes. Distillation attacks involve tricking models into revealing the chain of thought behind their outputs, which can then be used to train smaller models. For instance, an attacker cleverly framed a query as a translation request to extract specific cognitive patterns from Anthropic's Claude model. Such findings raise concerns over intellectual property security and competitive dynamics in the AI landscape, highlighting the urgent need for enhanced defenses against these tactics as the AI arms race intensifies.