Anthropic has officially unveiled a new defensive technology for its Claude models, aimed at curbing industrial-scale illegal knowledge distillation attacks. Early data shows that the newly deployed defense mechanisms have successfully reduced the amount of data stolen by attackers by 90% to 99%, addressing industry concerns about the low-cost replication of frontier model capabilities.
The core of this protection mechanism lies in restricting external access to Claude's reasoning process, reducing the number of reasoning tokens extractable by attackers to 1% to 10% of the original amount. The system maintains an extremely low false positive rate (<0.1%) while intercepting large-scale information extraction, with robustness comparable to biological weapon defenses. Additionally, Anthropic has banned tens of thousands of accounts suspected of distillation and publicly named Zhipu AI for attempting to distill capabilities from Claude during the development of GLM-5.3, raising national security risk concerns.
This move not only significantly increases the cost of replicating frontier capabilities but also demonstrates viable alternatives against fraud account networks by publishing actionable defense solutions, opposing industry inaction. This progress has profound implications for maintaining large model ecosystem security, protecting intellectual property, and addressing potential national security risks.