After OpenAI, Anthropic also claimed that Chinese artificial intelligence (AI) corporations had illicitly extracted outputs from its AI models.
Anthropic said on the 23rd (local time) that it had confirmed that three Chinese AI corporations—DeepSeek, MoonshotAI and MiniMax—illegally extracted the capabilities of its AI model "Claude." Anthropic said, "Chinese corporations generated more than 16 million conversations with Claude through about 24,000 fake accounts," adding, "DeepSeek siphoned outputs through 150,000 conversations, MoonshotAI through 3.4 million, and MiniMax through 13 million."
Anthropic said the Chinese corporations used a "distillation" technique. Distillation is a method that uses the answers produced by another AI model as training data to create a model with similar performance. U.S. AI corporations also use distillation. That is because it can be used to build a lightweight sub-model with abilities comparable to a top-tier model. Google, for example, built "Gemini Flash" based on the higher-end model "Gemini Pro."
However, using the distillation technique on a competitor's model is effectively regarded as theft or exfiltration.
Anthropic warned that models extracted illegally are likely to lack necessary safety measures and could pose serious national security risks. U.S. AI models, including Claude, have systems to prevent threat actors from abusing AI for biological weapons development or malicious cyber activity, but models extracted through distillation are likely to have such safeguards removed, the company said.
Anthropic added that it had banned access to Claude within China for national security reasons, but Chinese corporations still circumvented this and used distillation. Anthropic noted, "We have consistently supported export controls to China (on AI chips) so the United States can maintain its competitive edge in AI, but distillation attacks undercut export controls by narrowing the U.S.-China AI gap in another way."
The company said, "To counter Chinese corporations' distillation attacks, we are strengthening safeguards by developing detection tools that can identify coordinated activity across large clusters of accounts and by tightening verification procedures," but emphasized that solving this problem will require cooperation among the AI industry, cloud providers and the government.
Earlier, Anthropic Chief Executive Officer Dario Amodei strongly criticized the Donald Trump administration's policy allowing exports of Nvidia's high-performance AI chips to China, saying it was "similar to selling nuclear weapons to North Korea" and "a mistake with enormous implications for national security."
Earlier, OpenAI, the developer of ChatGPT, also expressed concern in a memo submitted to the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party on the 12th that Chinese corporations, including DeepSeek, were extracting outputs from U.S. AI models by using distillation.