Anthropic Distillation Forensics: 16M Exchanges, 24K Fraudulent Accounts, Three Chinese Labs

Anthropic disclosed the most comprehensive model distillation attack forensics to date. MiniMax generated 13M+ exchanges targeting agentic coding and tool use. Moonshot AI targeted agentic reasoning and computer-use agents (3.4M exchanges). DeepSeek focused on reasoning and censorship-bypass generation (150K+ exchanges). Total: 16M exchanges across 24K fraudulent accounts, attributed via IP correlation, metadata, and cross-industry corroboration. DeepSeek notably sought help generating "censorship-safe alternatives" to politically sensitive queries. Anthropic's response includes behavioral fingerprinting classifiers and "response shaping" defenses. This is the first publicly documented case of state-adjacent industrial distillation at this scale. Anthropic Blog | The Hacker News