Anthropic Says New Safeguards Cut AI Model Distillation Attempts by Up to 99%

Anthropic's new safeguards aim to limit the extraction of its AI model's reasoning capabilities. This process, known as distillation, can allow other developers to train their own systems using Claude's outputs, potentially reducing the time and cost required to develop advanced AI. The company claims these measures have significantly cut the volume of reasoning tokens extracted, reducing it by "10 to 100 times lower than it was before."

The concern over model distillation arises when companies allegedly use another provider’s models in ways that violate access restrictions or terms of service. This practice could allow them to obtain capabilities that would otherwise require substantial investment. Anthropic specifically cited Chinese AI developer Z.ai, also known as Zhipu, as an example of a company it accused of attempting to extract Claude’s capabilities while developing its GLM-5.3 model.

Anthropic's actions, including banning tens of thousands of accounts and making it harder for attackers to use proxy networks, highlight the intensifying competition and intellectual property concerns in the AI sector. The company also warned that unauthorized transfer of advanced AI capabilities could create safety and national security risks by undermining safeguards developed by the original model provider.