# Anthropic says Chinese labs strip-mined Claude

> Anthropic says Alibaba, Moonshot and DeepSeek quietly pulled millions of Claude answers to train rival models.
- Title: Anthropic says Chinese labs strip-mined Claude · CNBC
- Summary: Anthropic says Alibaba, Moonshot and DeepSeek quietly pulled millions of Claude answers to train rival models. The cheapest way to build a frontier model may…
- Keywords: anthropic, china ai, distillation, claude, ai security, business, finance, Chinese, labs, strip-mined, CNBC
- Source: CNBC — https://cnbc.com/2026/09/11/chinese-ai-labs-moonshot-deepseek-alibaba-anthropic.html
- Author: Jenny Lee
- Published: 2026-09-11T00:48:15+00:00
- Read time: 1 min
- Topics: anthropic, china ai, distillation, claude, ai security, business, finance
## Anthropic names names
In a threat report out Thursday, Anthropic said it caught and shut down large-scale efforts by Alibaba, Moonshot and DeepSeek to train their models on Claude.
## What illicit distillation means
You pump a stronger model for outputs, then feed those outputs to your own model so it copies the behavior. Doing it without permission breaks the terms of service.

> outputs from a more capable AI model are used to train another model and replicate some of its capabilities without authorization
## 151 million exchanges
Anthropic calls the Alibaba campaign the biggest it has ever measured: 151 million exchanges between May and July, peaking near 3 million a day from 3,500-plus fake accounts.
## Kimi users got Claude
Moonshot quietly forwarded some Kimi customer requests to Claude and showed Claude's answers back as Kimi's. In one 10-day stretch that was nearly 300,000 requests.
## Reasoning traces harvested
Anthropic says Moonshot kept some of those exchanges and pulled out Claude's reasoning transcripts as training data. Total attributed to Moonshot: over 23 million exchanges.
## Private data went along
Some relayed requests carried sensitive information from individuals, multinationals and state-linked actors. DeepSeek ran a similar relay, 12 million attacks in 14 days of July 2026.

> These practices are likely inconsistent with privacy laws and the labs' own terms of service
## Key takeaway

The cheapest way to build a frontier model may be to secretly query someone else's, and the host can see it happening.