Anthropic anti-distillation cuts Claude extraction 90–99%
Anthropic says new safeguards cut reasoning-token extraction from Claude by 10–100x, relying on preserved thinking, a two-stage classifier and fast account bans, including one network of 20,000+ fraudulent accounts.
New safeguards at @AnthropicAI cut the data distillers' harvest from Claude by about 90 to 99%, early evidence suggests. Measured on its own platform, reasoning tokens extracted fell 10 to 100x after the classifier went in Preserved thinking, a two-stage classifier, and fast account bans are key parts of the stack One network ran more than 20,000 fraudulent accounts at once Zhipu and Moonshot are named as past distillation cases The company says distillation lets rivals mimic frontier capability for a fraction of the development cost. I would count extraction defenses as part of a lab's product, alongside the model itself.