Anthropic Disrupts Industrial-Scale Claude Distillation Attacks by Seven China-Based AI Labs

Anthropic has revealed that it identified and disrupted a coordinated campaign of illicit knowledge distillation targeting its Claude AI model. The attacks, attributed to seven AI laboratories based in China, were described as 'industrial-scale' and involved systematically querying Claude to extract outputs that could be used to train competing models. The named entities include Alibaba, Moonshot, DeepSeek, Z.ai (also known as Zhipu), and MiniMax.

Knowledge distillation is a standard machine learning technique where a large 'teacher' model guides the training of a smaller 'student' model. However, Anthropic asserts that the scale and method of these operations crossed into unauthorized territory, effectively appropriating Claude's capabilities without proper licensing. The company has since taken measures to disrupt the attacks, though specific technical countermeasures were not disclosed.

The disclosure highlights growing tensions in the AI industry over intellectual property and the misuse of API access. As AI models become more powerful, the risk of distillation attacks—where competitors extract knowledge to accelerate their own development—poses a significant threat to the competitive advantage of frontier AI developers.

Anthropic's action underscores the need for robust monitoring and rate-limiting of API queries, as well as legal and technical safeguards against unauthorized model extraction. The company has not released a software patch, as the issue is primarily a terms-of-service and access control matter rather than a traditional vulnerability.

While Anthropic's attribution is firm, independent verification of the specific labs' involvement is pending. The broader implications for AI security and the ethical boundaries of model training are likely to fuel further debate and regulatory scrutiny.