Anthropic says new safeguards sharply reduce AI distillation attacks

Anthropic's reported success in sharply reducing distillation attacks highlights a growing concern for leading AI developers. The practice, which involves copying advanced AI capabilities, is seen by companies like Anthropic, OpenAI, and Google as enabling rivals to reproduce sophisticated models more quickly and at a lower cost. This issue gains particular urgency amid intensifying US-China competition in the AI sector, where firms aim to protect proprietary model development.

The new safeguards primarily target Claude's reasoning processes, restricting access and making it harder for attackers to use proxy networks for repeated probes. Anthropic claims its safety system is "roughly as robust as those that prevent Claude from helping build bioweapons" and has "a false positive rate of less than 0.1%". This suggests a technical approach focused on making the extraction of core AI logic significantly more difficult, rather than simply blocking access.

By identifying and banning tens of thousands of accounts linked to these campaigns, Anthropic demonstrates an active enforcement strategy. The specific mention of Z.ai (Zhipu) and its GLM-5.3 model development illustrates the direct competitive implications. Anthropic warned such practices pose safety and national security risks, indicating that AI firms are prepared to attribute alleged distillation efforts publicly.