We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·Jun 25·5 min read
Anthropic: Alibaba-Linked Operators Used 25k Accounts to Mine Claude for Qwen
TL;DR
Anthropic identified a significant issue where operators linked to Alibaba allegedly used 25,000 fake accounts to extract capabilities from their AI model, Claude. This operation, which occurred between April and June 2026, resulted in over 28.8 million interactions focused on software engineering and reasoning tasks.
✦ Why It Matters
Engineers should be aware of the risks of data harvesting and the importance of securing AI model outputs.
Key Takeaways
- Implement stricter verification processes for cloud accounts.
- Monitor interactions with AI models to detect unusual patterns.
- Consider the implications of data harvesting on model training.
- Engage in policy discussions about AI access and competition.
Full Summary
Anthropic, the creator of the AI model Claude, has accused operators associated with Alibaba's Qwen AI lab of exploiting nearly 25,000 fraudulent accounts to mine Claude's capabilities. This campaign, which took place from April 22 to June 5, 2026, generated an astonishing 28.8 million exchanges, primarily targeting software engineering and agentic reasoning—key functionalities that enhance Claude's utility for developers.
The methodology involved using these accounts to interact with Claude, effectively harvesting its outputs for competitive advantage. This situation has escalated into a broader policy debate regarding access to advanced AI models and the verification of cloud accounts.
Anthropic argues that the frontier AI race now hinges on preventing rivals from using high-quality model outputs as training data. The implications of this incident could reshape how AI companies approach data security and competitive practices.