Strengthening our defenses against distillation attacks

Anthropic Research has announced that its latest security measures have reduced the effectiveness of illicit distillation attacks on its Claude models by more than 90%. Distillation, a covert industrial‑scale campaign to extract a model’s capabilities and replicate them in another model, has been used to help foreign laboratories keep pace with U.S. companies. The company’s new defenses focus on limiting access to Claude’s internal reasoning transcripts, which are the most valuable data for distillers. Early evidence indicates that the amount of data harvested by distillers has fallen between 90 and 99 percent. Anthropic’s strategy includes controlling where reasoning can be shared, preventing new API accounts from editing prior context, and making reasoning transcripts accessible only to the organization that generated them. The company also deployed a two‑stage classifier system to detect and block extraction attempts, mirroring safeguards used for cybersecurity and biology. While the measures have caused some revenue loss and required legitimate users to update integrations, Anthropic believes the trade‑offs are justified to protect frontier AI capabilities and mitigate national security risks. The company hopes its post will help other developers understand how to defend against distillation and the associated costs.