Anthropic just exposed industrial-scale AI model theft by the Chinese labs behind DeepSeek, Moonshot AI, and MiniMax.

They ran massive distillation campaigns targeting Claude's capabilities. We're talking over 16 million queries across ~24,000 fraudulent accounts.

My take:

  1. ROI on a distillation attack is mind-blowing, so all three major labs GDM, OpenAI, and Anthropic are targets.
  2. Bulk accounts behind a proxy are at the heart of the distillation system.
  3. Distillation is an attack that is challenging to protect against. Attackers are becoming more creative in designing their queries, so it's challenging to distinguish distillation traffic with high precision. And sometimes these accounts may even be paying you!
  4. OpenAI and Anthropic are leading the pack, being vocal about their counter-efforts to protect themselves from being accused of insufficient enforcement of export controls.