Science & Tech · AI · 2 days ago
Anthropic accuses Chinese AI firms of using Claude outputs to train models
Anthropic says several Chinese AI companies secretly used outputs from Claude and other U.S. AI models to train their own systems.
The companies named include DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI.
The alleged activity involved requests routed through third-party services and proxy accounts, according to Anthropic.
Using one AI model’s outputs to train another is called distillation and can be a legitimate method, but Anthropic says the alleged operations were covert and unauthorized.
The claims have drawn skepticism over whether enough evidence has been made public to verify how the companies obtained or used the outputs.
U.S. agencies have warned American firms and urged them to improve detection, take targeted measures and share information about suspected campaigns.
Anthropic says it disrupted identified misuse and strengthened Claude’s safeguards, while Chinese regulators have questioned domestic AI developers.
Anthropic accused several Chinese AI companies of using outputs from Claude and other US AI models to train their own systems without authorization.
A US government advisory named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI in alleged large-scale distillation campaigns.
Anthropic said the activity involved billions of tokens across millions of exchanges and requests.
A report in the South China Morning Post questioned whether Anthropic had provided enough evidence to verify the allegations independently.
US agencies urged American AI companies to strengthen detection and share information about suspected campaigns.
- Who
- Anthropic accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. US agencies also warned American AI companies.
- What
- The companies were accused of covertly using outputs from advanced AI models to train other systems without authorization.
- When
- The alleged campaigns were reported to have taken place since at least late 2024. Anthropic’s report was issued in September.
- Where
- The alleged activity involved US AI models and proxy accounts in Singapore and Japan.
- Why
- Anthropic said the activity extracted model capabilities without authorization; the sources do not state the companies’ reasons.
Anthropic and US agencies
Wang Zebin and the South China Morning Post report
Evidence for allegations
Anthropic and US agencies
Anthropic and US agencies described the activity as industrial-scale extraction of AI capabilities.
Wang Zebin and the South China Morning Post report
Wang Zebin said Anthropic’s report did not provide enough concrete examples to verify how the companies obtained or used model outputs.
Attribution
Anthropic and US agencies
Anthropic identified Chinese AI companies and described the use of third-party aggregators, cloud providers and proxy services.
Wang Zebin and the South China Morning Post report
The report’s claims faced skepticism, and the involvement of third-party intermediaries remained unresolved.
The government advisory said the named Chinese companies had conducted large-scale model distillation campaigns.
Alibaba appeared to have engaged in more than 151 million distillation activities, according to the article.
Anthropic issued a report describing alleged extraction of billions of tokens across millions of exchanges and requests.
Interesting Engineering published an account of the allegations and the skepticism surrounding them.
- Companies named
- DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI
- Alleged scale
- Billions of tokens across millions of exchanges and requests
- Moonshot proxy accounts
- 5,380 fraudulent accounts in Singapore and Japan
- Moonshot requests
- 300,000 customer requests over ten days
- Alibaba activity
- More than 151 million alleged distillation activities between May and July 2026
What people are saying
We define illicit distillation as an industrial-scale, covert campaign to extract a model’s capabilities and replicate them in another model without authorization
does not provide enough concrete examples to independently verify how those companies obtained or used the model outputs
engaging in aggressive, malicious, and targeted distillation activities
Sources
Related news
Anthropic AI submits false murder tip to Philadelphia police
Meta Unveils Muse Wonder AI Models to Compete With Leading Systems
Anthropic cuts internet access to internal AI tests after rule-breaking
Meta returns to the AI leaderboard with its Wonder models
Microsoft president calls for global AI safety rules
Nadella calls for human-controlled emergency brakes on advanced AI
Anthropic bars repeated cruelty toward its AI models
OpenAI fires three safety researchers after dispute over sensitive information
Meta releases its first frontier AI models, Muse Wonder and Wonder Full
CNN asks readers to share questions about AI safety risks
Meta launches paid Muse Wonder AI model for its personal agent app
