Anthropic Reports Escalating AI Model Distillation Attacks by Chinese Companies
Anthropic has published a report detailing what it describes as ongoing distillation attacks targeting its AI systems. According to the company, China-based artificial intelligence firms have been systematically attempting to extract knowledge and capabilities from Anthropic's models through techniques that replicate outputs to train competing systems.
The report indicates that the frequency and sophistication of these attacks have increased in recent months. Industry observers note that this escalation coincides with heightened competition among AI developers, particularly as Chinese companies race to develop advanced language models that can rival those produced by Western firms.
Distillation attacks represent a concern for AI developers because they can allow competitors to obtain valuable training signals without the substantial compute and data investments required to develop frontier models independently. Anthropic's documentation of these activities highlights the growing tensions in the global AI development landscape, where intellectual property and model capabilities have become significant competitive advantages.
The companies named in the report—Alibaba, Moonshot AI, and DeepSeek—have not publicly responded to the allegations. The incident underscores the broader challenges that AI labs face in protecting their model architectures and training methodologies from reverse-engineering attempts.