Anthropic Distillation Detection: Industrial-Scale Model Theft

Anthropic disclosed distillation campaigns by three Chinese AI labs through 24,000 fraudulent accounts generating 16M+ exchanges. MiniMax drove 13M (code generation focus), Moonshot generated 3.4M (targeting agentic reasoning and tool use), DeepSeek produced 150K (reasoning and censorship-safe alternatives). Detection used behavioral fingerprinting, IP correlation, and chain-of-thought elicitation classifiers. MiniMax pivoted to new Claude models within 24 hours of release. For builders operating APIs, the detection patterns — coordinated timing, proxy service mixing, abnormal prompt structure — are directly applicable to protecting their own endpoints. (Anthropic Blog, The Hacker News)