TechCrunch Industry News Anthropic accuses Chinese AI labs of mining Claude as US debates AI chip exports
Feb 24, 2026
A rundown of allegations that Chinese labs used thousands of fake accounts to extract a leading AI model’s capabilities. A clear explanation of model distillation and how rivals can copy advanced systems. Fast-moving open-source projects claiming to match frontier models on the cheap. Debate over chip export controls and the security risks of poorly safeguarded distilled models.
AI Snips
Chapters
Transcript
Episode notes
Mass Distillation Attacks Target Claude Capabilities
- Anthropic alleges DeepSeek, Moonshot AI, and Minimax created 24,000+ fake accounts and generated over 16 million exchanges to distill Claude's capabilities.
- Attacks focused on agentic reasoning, tool use, and coding, with Minimax alone redirecting nearly half its traffic to siphon Claude on launch.
Chinese Labs Rapidly Release Competitive Models
- OpenAI earlier accused DeepSeek of distillation and DeepSeq released R1 and is set to release DeepSeq v4 claiming performance rivaling U.S. frontier labs.
- Moonshot released KimiK 2.5 and a coding agent last month while Minimax focused heavily on agentic coding.
Coordinate Defenses Across Industry And Policy
- Anthropic plans to invest in defenses that make distillation attacks harder to execute and easier to detect and calls for coordinated industry and policy responses.
- It urges collaboration across AI firms, cloud providers, and policymakers to counter extraction at scale.