Anthropic accuses three Chinese AI systems of stealing Claude

Anthropic takes on the bull by the horns. In a furious blog post, the creator of Claude explains having identified an « industrial-scale campaign » led by three Chinese labs, DeepSeek, Moonshot, and MiniMax. Their wrongdoing: illegally extracting the capabilities of Anthropic's AI model to improve their own models at lower costs.

When the distillers are distilled

The process, called "distillation," is not in itself a crime. It's even a common practice of training a smaller or less performant model from the responses of a more powerful model. Large AI companies use it to produce lighter, cheaper versions for their customers. Therefore, the method itself is not illegal.

What poses a problem here, according to Anthropic, is the clandestine use: instead of distilling their own models, these labs would have used Claude as an "involuntary teacher." This would allow them to acquire certain capabilities "in a fraction of the time and cost " necessary to develop them independently. According to the American startup, these three companies would have created over 24,000 fraudulent accounts and generated over 16 million interactions with Claude.

DeepSeek would have used only 150,000 interactions with Claude, compared to 3.4 million for Moonshot and 13 million for MiniMax. Anthropic is not alone on this battlefield: OpenAI complained directly to the White House, accusing DeepSeek of using the same tactic, this time with GPT.

The three Chinese labs would have thus extracted advanced reasoning capabilities, alternatives to politically sensitive queries, and programming for agent coding. The volume and repetition of queries would have betrayed a logic of systematic extraction rather than normal use.

For Anthropic, this practice originating from Chinese competitors represents a national security risk for the United States. American models include safeguards to limit certain uses (sensitive biology, cyberattacks, etc.). If a model is illicitly distilled, these protections may not be replicated, while the raw capabilities remain. Military or surveillance systems could take advantage of this.

The irony is that these same companies continue to defend in court their right to train their models with oceans of protected content, often without permission or payment to the authors. We have seen this again with Mistral. On one hand, distillation would be an intolerable infringement on innovation; on the other hand, the massive collection of text, images, and data would be considered fair use.

Regardless, Anthropic does not want to sit idly by. The startup is investing in defense mechanisms to curb illegal distillation campaigns: detection and identification systems, tracking coordinated activities across thousands of accounts, sharing technical indicators with other AI companies, strengthening verification procedures, etc. But the company acknowledges that no single actor can solve the problem alone: at this scale, the response will have to be coordinated among industrialists, hosts, and public decision-makers.

👉🏻 Follow tech news in real-time: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source : Anthropic