Anthropic says new safeguards cut AI model-distillation attacks by 90–99%

What happened

Anthropic said on Monday that new safeguards designed to stop rivals copying the capabilities of its Claude models have cut data harvested in so-called distillation attacks by about 90% to 99%. The measures restrict access to Claude's reasoning, make it harder to use proxy networks to repeatedly probe protected reasoning, block extraction attempts, and have banned tens of thousands of accounts linked to distillation campaigns. Anthropic named Chinese AI startup Z.ai (Zhipu) as an example of a company that attempted to distill capabilities from Claude while developing its GLM-5.3 model, warning the practice poses safety and national security risks.

Why it may matter to you

Model distillation — training new models on another lab's model outputs — is an AI-use-of-work issue aimed at the labs' own outputs: unauthorised extraction of the training value in a model. Anthropic's reported 90–99% reduction, if accurate, would materially raise the cost of reproducing advanced capabilities and could become the industry's reference anti-distillation playbook.

What is not yet known

Independent verification of the 90–99% reduction claim, which comes from Anthropic itself. Whether other labs will adopt comparable measures, and whether regulators or courts treat model distillation as unlawful use of a competitor's outputs.