Chinese AI firms used Claude to train their AI, Anthropic says.

Chinese AI companies DeepSeek, Moonshot, and MiniMax reportedly used the chatbot Claude to acquire capabilities for their own models through a process called distillation, according to Anthropic.

The company said these firms created over 16 million interactions with Claude using around 24,000 fake accounts, violating Anthropic’s terms of service and regional access restrictions.

Anthropic described distillation as training a less-capable model on outputs from a stronger one.

The company warned that such illicit activities could pose national security risks if these models are open-sourced, as their capabilities could spread beyond control.

Anthropic highlighted that these campaigns are increasing in complexity and support the case for export controls on chips, which could limit access to model training resources.

The companies targeted reasoning tasks, censorship-safe responses, coding, and data analysis, with Anthropic noting that MiniMax shifted traffic to its latest model within 24 hours of the company’s release.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

The alleged Claude distillation campaign was described as an industrial-scale operation

Anthropic called it an “industrial-scale” effort; it said DeepSeek, Moonshot, and MiniMax ran about 24,000 fake accounts that produced over 16 million Claude interactions, violating terms of service and regional access limits 1.
OpenAI earlier told lawmakers that Chinese AI firm DeepSeek sought to copy models from the ChatGPT maker plus other leading U.S. AI labs for its own training 1.
Anthropic said it is sharing threat intelligence with other AI companies and tightening access controls to coordinate defenses against distillation attacks 2.
Some critics called the claims a lobbying play; they said training on rivals’ outputs happens across the industry, which muddies the ethics 3.

Distillation challenges chip export controls and is driving a new AI security push

Anthropic said chip export controls miss part of the problem because limits on GPUs curb direct training while distillation can still pull capabilities from model outputs 4.
The episode treats model outputs as intellectual property that can be stolen, which raises the need to guard behavior plus access routes 5.
AI labs are moving toward a wider security set that includes detection tools, shared intel, and stricter access gates for automated capability extraction 2.
Anthropic warned that distilled models may drop safety guardrails; open-sourcing would spread those capabilities beyond any one company or government 1.