Anthropic Says New Safeguards Sharply Reduce AI Distillation Attacks

3-line summary

• Anthropic said new safeguards restricting access to Claude's reasoning process and making it harder to repeatedly probe its protected reasoning through proxy networks have sharply reduced "AI distillation" attacks, in which rivals mass-query a model to harvest its outputs and reproduce its capabilities faster and more cheaply, according to Reuters. The company said the detection system has a false-positive rate below 0.1%.
• Anthropic said the safeguards cut data harvested through distillation attempts by roughly 90-99% and reduced extracted reasoning tokens by 10-100x, and it blocked tens of thousands of accounts linked to distillation campaigns. The company cited Chinese AI startup Z.ai (Zhipu) as an example, alleging the firm attempted to distill Claude's capabilities while developing its GLM-5.3 model; Z.ai did not immediately respond to a request for comment.

This summary is our own processing of the source article; see the source link below for full context.