Key facts
- Anthropic alleged Alibaba-affiliated operators conducted a large-scale AI model distillation campaign against its Claude chatbot.
- The campaign allegedly involved over 28.8 million exchanges using nearly 25,000 fraudulent accounts between April 22 and June 5.
- Anthropic claims the operations targeted Claude's agentic reasoning, software engineering, and long-horizon planning capabilities.
- The company urged Congress to strengthen export controls, expand intelligence sharing, and penalize firms for large-scale model extraction.
- Anthropic views large-scale model distillation as a national security issue that could accelerate China's AI capabilities.
Anthropic has called on Congress to implement stronger protections against AI model distillation, alleging that operators affiliated with Alibaba conducted the largest known campaign to extract capabilities from its Claude chatbot. In a letter dated June 10, Anthropic detailed that these operations involved over 28.8 million exchanges with Claude between April 22 and June 5, utilizing approximately 25,000 fraudulent accounts. The company stated that these 'distillation attacks' targeted Claude's advanced reasoning and planning abilities, enabling competitors to replicate sophisticated AI behavior without the substantial costs of training frontier models.
Anthropic emphasized the brazen nature of the campaign, noting Alibaba's listing on the New York Stock Exchange and its business operations in the United States. The AI firm framed large-scale model distillation not just as an intellectual property concern but as a national security issue that could bolster China's military and cyber AI capabilities while diminishing the United States' technological advantage. Anthropic argued that such practices effectively subsidize competitors by allowing them to capture the returns on American AI investments without bearing the associated research, development, and compute costs.
This appeal to Congress comes as the U.S. government intensifies efforts to safeguard its AI leadership. President Donald Trump recently signed an executive order aimed at enhancing AI-powered cybersecurity. Anthropic proposed several measures for lawmakers to consider, including expanding intelligence sharing between AI developers and the government, clarifying antitrust rules to facilitate information sharing on distillation attacks, strengthening export controls on advanced AI chips and compute power, closing loopholes for Chinese firms accessing overseas data centers, and imposing penalties on entities involved in large-scale model extraction.
This is not the first time Anthropic has raised concerns about Chinese AI developers. In February, the company alleged that DeepSeek, Moonshot AI, and MiniMax engaged in similar distillation activities. While some observers have noted that AI companies often use similar techniques for training their own models, Anthropic distinguishes between legitimate model distillation for creating smaller, more efficient models and unauthorized extraction of frontier model capabilities through fraudulent access, which violates their terms of service. The broader debate around AI distillation was further complicated by Elon Musk's testimony regarding xAI's use of OpenAI models in training Grok, highlighting the industry's ongoing discussion about the line between legitimate training and unauthorized extraction.