Anthropic AI Theft: Chinese Firms Accused of Distilling Claude

Anthropic has accused Chinese giants DeepSeek, Moonshot, and MiniMax of a massive theft of the Claude model's intelligence. Through millions of interactions, the Chinese firms attempted to "distill" the capabilities of the American AI, sparking global security concerns.

The War for AI Supremacy Intensifies

The company Anthropic, recently valued at $380 billion, has taken up arms against Chinese competition. It has accused a trio of technology firms developing artificial intelligence of so-called "industrial distillation." The companies DeepSeek, Moonshot, and MiniMax allegedly exploited millions of interactions with the Claude model to leapfrog the development of their own systems. According to the American giant, this is not just theft, but a global-scale security risk.

It is the scenario that Western developers feared most. Instead of lengthy and costly in-house development, the competition simply "borrows" the brain of your best model to teach their own how to think. According to Anthropic's Monday statement, this is exactly what happened.

In a detailed report, the firm revealed that three Chinese AI companies launched a massive coordinated campaign. During this operation, over 24,000 fake accounts generated more than 16 million interactions with the Claude model. The goal was singular: to "distill" its capabilities.

What is "distillation"?
In technical jargon, this is a process where a smaller and "simpler" model learns from the outputs of a more powerful model (in this case, Claude). Instead of expensive training on raw data, the Chinese models learned directly from the finished, high-quality responses of the American AI.

Security Risks and Theft of Know-how

“These campaigns are growing in intensity and sophistication. The window for response is narrowing, and the threat transcends any single company,” Anthropic warns in its blog post.

The problem, however, is not just the theft of intellectual property. Anthropic points out that such "stolen" models often lack critical safety guardrails. If these unregulated but highly capable systems were to enter the open-source space (as often happens with DeepSeek), the security risks could spiral out of the control of any government.

A Game of Cat and Mouse: Pivoting Within 24 Hours

The report reveals fascinating details about how the entire operation was conducted. Each of the Chinese firms specialized in a different area of development:

DeepSeek: Targeted logical reasoning capabilities and creating versions that bypass censorship and handle politically sensitive topics.
Moonshot: Focused on "agentic" reasoning, tool use, coding, and data analysis.
MiniMax: Focused on the orchestration of complex tasks.

The Audacity of the Attackers Knows No Bounds
The incident with MiniMax illustrates the speed of adaptation. “When we released a new model during their active campaign, they reacted within 24 hours. They immediately redirected nearly half of their traffic to our new system to capture its latest capabilities,” Anthropic states.

The Argument for Stricter Chip Controls

Anthropic, which recently closed a $30 billion investment round, is using this incident as a strong argument for stricter chip export controls.

The logic is relentless: If you restrict access to the most powerful hardware, you hinder not only direct model training but also the efficiency of this "distillation." Without top-tier chips, it is much harder to efficiently process and implement data obtained from foreign models.

Context: Not the First Time

The current allegations come just weeks after a similar warning was issued by OpenAI. That company warned US lawmakers that DeepSeek is specifically copying the architecture of ChatGPT.

It appears that the Chinese AI sector, facing hardware sanctions, has found a way to keep pace—not through raw computing power, but by "parasitizing" Western innovation. None of the named Chinese companies have yet responded to Reuters' request for comment.