Anthropic accused three Chinese AI companies on Monday of using 24,000 fake accounts to extract capabilities from its Claude model. DeepSeek, Moonshot AI, and MiniMax generated more than 16 million exchanges with Claude to train their own models.
The technique, called distillation, lets smaller AI models learn from more powerful ones by studying their outputs. AI companies routinely use distillation on their own models to create cheaper versions. But Anthropic says these Chinese labs used it to copy capabilities that took years and hundreds of millions of dollars to develop.
How the Extraction Worked
The three companies created fraudulent accounts and used proxy services to access Claude despite regional restrictions. DeepSeek generated 150,000 exchanges focused on reasoning tasks and censorship-safe responses. Moonshot AI created 3.4 million exchanges targeting coding and tool use. MiniMax ran 13 million exchanges, redirecting half its traffic to Claude’s latest model within 24 hours of its release.
Anthropic tracked the campaigns through IP addresses, request patterns, and infrastructure indicators. The prompts showed unusual patterns. DeepSeek asked Claude to explain its internal reasoning step by step, generating training data at industrial scale.
National Security Concerns
Anthropic warned that distilled models lack safety guardrails built into American AI systems. The company builds protections to prevent misuse for bioweapons development or cyber attacks. Models trained through distillation strip out these safeguards.
The accusations come as U.S. officials debate AI chip export controls aimed at limiting China’s AI development. Anthropic argued that the scale of these extraction campaigns requires access to advanced chips, reinforcing the case for export restrictions.
OpenAI made similar accusations against DeepSeek earlier this month, telling House lawmakers the company used distillation to mimic its products. DeepSeek, MiniMax, and Moonshot have not commented on the allegations.