3 hrs ago
Anthropic Accuses Chinese AI Labs of Secret Claude Data Routing
Anthropic says some Chinese AI companies secretly sent their customers’ questions to Claude, an Anthropic chatbot.
The companies allegedly used fake accounts to get around Anthropic’s restrictions on access from China.
Moonshot AI was accused of sending hundreds of thousands of requests this way.
Anthropic says the companies used Claude’s answers to help train their own AI systems.
Some of the redirected conversations reportedly contained private or sensitive information.
One example involved a question about CCTV tracking in China.
Another reportedly exposed credentials for a Russian government database.
Anthropic says it stopped the schemes it found and is watching for more activity.
Anthropic accused Moonshot AI, DeepSeek and other Chinese laboratories of routing customer prompts through Claude using fraudulent accounts.
Anthropic said Moonshot used 5,380 accounts to relay nearly 300,000 requests in one 10-day period.
The company said Moonshot generated more than 23 million exchanges between May and July for alleged model distillation.
Anthropic reported that some redirected traffic contained sensitive information, including database credentials and police-case material.
The company said it disrupted the identified schemes and alleged the practices may violate privacy laws and terms of service.
- Who
- Anthropic accused Moonshot AI, DeepSeek, Xiaomi and other Chinese AI laboratories of the alleged activity.
- What
- The laboratories were accused of routing customer conversations through Anthropic’s Claude models using fraudulent accounts and using the responses for model training.
- Where
- The alleged account networks appeared largely based in Singapore and Japan, while the activity involved Chinese AI companies and users’ traffic reaching Anthropic’s servers.
- When
- Anthropic said Moonshot’s reported activity occurred between May and July; its earlier allegations were raised in February, and the latest report was issued Thursday.
- Why
- Anthropic said the activity was intended to bypass access restrictions and obtain Claude outputs for training competing AI systems, a practice it calls illicit distillation.
Key facts
- Company making allegations
- Anthropic
- Main companies named
- Moonshot AI, DeepSeek, Alibaba, Zhipu, SenseTime, MiniMax and Xiaomi
- Fraudulent accounts attributed to Moonshot
- 5,380 accounts, most appearing to be based in Singapore and Japan
- Moonshot traffic in one 10-day period
- Nearly 300,000 customer requests
- Moonshot exchanges between May and July
- More than 23 million exchanges
- Earlier figures cited by Anthropic
- More than 16 million exchanges using roughly 24,000 fraudulent accounts involving DeepSeek, MiniMax and Moonshot
- Sensitive information reportedly exposed
- Live credentials for a Russian government database, Chinese police-case material and a CCTV-related query
Quotes
Anthropic
The U.S.-based artificial intelligence company that issued the threat intelligence report
“raise concerns about the misuse of user data by PRC AI labs”
freepressjournal.in
Sources
Related news
Anthropic Disrupts Alleged Russian and Chinese AI Campaigns
Anthropic Warns Claude Users of Malware-Stealing Login Sessions
DeepSeek Helps Chinese Hackers Scale Up Cyberattacks, Researchers Say
Anthropic Resumes External AI Cyber Tests After Claude Incidents
Anthropic Discloses Claude Breach After Accidental Internet Access
Japanese Money Repatriation Could Pressure Global Risk Assets
Sujeet’s Asian Games Gold Path Leads Through Japan
Crude Surge and Asian Selloff Signal Weak Indian Opening
Gurindervir Singh Turns Commonwealth Setback Into Asian Games Focus
