Anthropic accuses DeepSeek and other Chinese rivals of mass data theft
10 hours ago
- #Intellectual Property Theft
- #AI Security
- #US-China Tech Rivalry
- Anthropic uncovered campaigns by three Chinese AI firms (DeepSeek, Moonshot AI, MiniMax) to illicitly extract capabilities from its Claude chatbot using 'distillation'.
- Distillation involves using outputs from a more powerful AI system to boost a less capable one, a practice common in AI development.
- The Chinese firms allegedly conducted 16 million exchanges with Claude via 24,000 fake accounts to siphon off capabilities at a fraction of the cost.
- Anthropic warned that illicit distillation bypasses U.S. export controls and poses national security risks, as safety guardrails may be omitted.
- OpenAI also accused Chinese firms of similar practices, calling it an effort to 'free-ride' on U.S. AI advancements.
- MiniMax was the largest offender, generating over 13 million exchanges, focusing on coding, reasoning, and tool use—Claude's strengths.
- Chinese firms circumvented China’s commercial access ban by routing traffic through proxy services and fraudulent accounts.
- Anthropic urged coordinated industry and government action to address the issue, which it says no single company can tackle alone.