Anthropic has accused the Chinese tech giant **Alibaba** of conducting an unprecedented model distillation operation by siphoning data from its AI, Claude. To bypass standard usage limits and harvest high-quality outputs, operators allegedly utilized approximately **25,000 fraudulent accounts** to generate over **28.8 million interactions**. This massive data collection effort was reportedly focused on extracting specialized knowledge in software engineering and logical reasoning to enhance Alibaba's own Qwen model series.
The scale of this operation marks it as the largest attack of its kind documented to date. Model distillation — the process of using a superior AI's responses to train a smaller or competing model — is a contentious practice in the industry. In this instance, Anthropic claims the sheer volume of synthetic data generated via automation suggests a deliberate, industrial-scale attempt to replicate Claude's proprietary capabilities.
- 25,000 accounts: Total number of fraudulent profiles used in the operation.
- 28.8 million: Specific number of interactions logged during the scraping phase.
- Core Focus: Software engineering and complex reasoning tasks.
