This page may contain stale information. Last updated: 2026-06-25

Definition

Model distillation extracts outputs from a stronger AI model to train or improve a weaker model. When conducted at industrial scale without authorization (“distillation attacks”), it enables competitors to replicate frontier capabilities without incurring full training costs.

Key Points

  • anthropic June 2026: alleged alibaba/qwen campaign — 25K fraudulent accounts, 28.8M Claude exchanges
  • Targets: software engineering and agentic reasoning capabilities
  • Exceeds combined prior DeepSeek/MiniMax/Moonshot campaigns per Anthropic
  • Congressional penalty and industry-wide defense coordination requested

Scale figures from Anthropic's letter; Alibaba has not publicly responded.

Sources