Anthropic says Alibaba used 25k accounts to mine Claude
Anthropic has accused Alibaba of using 25,000 accounts to systematically extract data from its Claude AI model, allegedly violating U.S. export controls under the Trump administration to steal proprietary capabilities. The company claims the Chinese tech giant conducted a large-scale operation to mine Claude for competitive advantage.
Background
- **Anthropic** is the AI company behind Claude, a major large-language model (LLM) competitor to OpenAI's ChatGPT and Google's Gemini. It was co-founded by former OpenAI employees.
- **Alibaba** is China's largest e-commerce and cloud-computing company. It has its own AI lab and LLM called Qwen, which competes directly with Western models.
- Anthropic is accusing Alibaba of running a large-scale "model extraction" attack: using ~25,000 fake accounts to query Claude, then using the outputs to train or copy capabilities for Alibaba's Qwen model. This is a known tactic in AI security — "distillation attacks" where an adversary probes a model to reverse-engineer its behavior.
- The article claims Alibaba did this after the Trump administration imposed tighter restrictions on Chinese AI companies accessing US AI services, which is the "defied Trump" angle in the headline.
- This matters because model extraction is increasingly central to the US-China tech rivalry. If proven, it would be one of the largest publicly alleged thefts of AI intellectual property, raising questions about how to protect frontier AI models from determined state-backed adversaries.