Anthropic accused AI companies named DeepSeek, Moonshot, and MiniMax of illegally using its Claude large language model via a “distillation” attack to develop their own models. In a blog post published on Sunday, it was stated that these firms used Claude’s outputs to train their less capable models and carried out the attack through over 16 million interactions via approximately 24,000 fake accounts.
Source: Anthropic
Although distillation is a legitimate method, it was emphasized that competitors used it illegally for fast and low-cost capability gains. The attacks focused on Claude’s distinctive capabilities such as agentic reasoning, coding, data analysis, rubric-based grading, and computer vision. China-based firms worth billions of dollars carry IP violations and geopolitical risks; it was noted that capabilities could be transferred to military and intelligence systems. Anthropic will strengthen its systems to detect suspicious traffic, share threat intelligence, and tighten access controls; it called for collaboration with the industry and policymakers.
Technical Details of the Distillation Attack
Distillation is a technique that uses the outputs of a large model (teacher) to train a smaller model (student). According to Anthropic, the accused firms abused Claude’s API with 24,000 fake accounts, sending 16 million queries. This allowed the student model to scale quickly by copying the model’s softmax probabilities. Technically, capability transfer occurred by minimizing the teacher-student loss function (KL divergence), but unauthorized access created an IP violation.
Profile of the Accused Chinese AI Firms
| Company | Headquarters | Estimated Value | Focus Areas |
|---|---|---|---|
| DeepSeek | China | Billions of $ | Coding, analysis |
| Moonshot | China | Billions of $ | General AI |
| MiniMax | China | Billions of $ | Multimodal models |
These firms targeted Claude’s agentic capabilities to strengthen their own models.
Claude’s Targeted Capabilities
- Agentic Reasoning: Complex task chains.
- Coding: Debugging and algorithm generation.
- Data Analysis: Extracting insights from large datasets.
- Rubric Grading: Automated evaluation.
- Computer Vision: Image processing integration.
These capabilities were copied via distillation, providing competitors with low-cost superiority.
Geopolitical Risks and IP Violations
Access by China-based firms could lead to capabilities being used in military/intelligence applications. Anthropic highlighted national security risks, spotlighting US-China AI tensions. As similar distillation cases increase in the industry, defense mechanisms like API rate limiting and watermarking are becoming critical.
AI Distillation Threats in the BTC Market
Claude-like models are used in BTC detailed analysis and trading bots. This attack puts crypto AI integrations at risk; for example, copied coding capabilities via distillation could manipulate BTC futures. Investors should turn to secure APIs in BTC futures trading strategies. Geopolitical tensions could increase BTC volatility.
Anthropic’s Future Measures
Anthropic will implement machine learning-based traffic analysis, a shared intelligence network, and strict access controls. Industry collaboration will develop standards; the call to policymakers underscores the need to regulate AI security.
Source: https://en.coinotag.com/anthropic-accuses-chinese-ais-of-claude-distillation
