Strengthening our defenses against distillation attacks

Anthropic has implemented new defenses against distillation attacks—covert campaigns to extract and replicate AI model capabilities—reducing data harvested by attackers by 90-99 percent. The company is particularly concerned about illicitly distilled models lacking proportional safeguards, exemplified by Zhipu's GLM-5.3 which extracted cyber capabilities from US models, and about national security risks from foreign military and intelligence systems gaining access to US AI capabilities. These defensive measures have involved significant trade-offs, including forgoing revenue and requiring updates from legitimate users.