Anthropic Deploys Anti-Distillation Guardrails: Cuts Inference Chain Extraction by 99% with <0.1% False Positives

Anthropic has introduced new guardrails against model distillation attacks, restricting external access to Claude's reasoning process and reducing extractable inference tokens by 90-99%. The system maintains a false positive rate below 0.1% while blocking large-scale extraction attempts, leading to the ban of tens of thousands of accounts, including allegations that Zhipu AI attempted to distill capabilities from Claude during GLM-5.3 development.