Anthropic launches new “anti-plagiarism” measures, says it has reduced distillation attack data by more than 90%

The Zhitong Finance App learned that Anthropic said on Monday that new measures aimed at preventing competitors from copying Claude's AI model capabilities have reduced the amount of data stolen in so-called “distillation attacks” by about 90% to 99%.

Model distillation has become a key concern for leading AI companies due to increased AI competition, and companies including Anthropic, OpenAI, and GOOGL.US (GOOGL.US) believe this practice enables competitors to reproduce advanced capabilities faster and at a lower cost.

According to Anthropic, protective measures limiting access to Claude's reasoning ability have drastically reduced extraction attempts, and “the number of inference tokens extracted by distillers” is now “10 to 100 times lower than before.”

The California-based company said it has made it more difficult, making it harder for others to use agent networks to repeatedly detect Claude's protected reasoning abilities, thereby blocking extraction attempts and impeding large-scale distillation activities.

According to Anthropic, its security systems are “roughly as robust as those that prevent Claude from assisting in the manufacture of biological weapons,” and that “the false alarm rate is less than 0.1%.”

Anthropic also previously blocked tens of thousands of accounts linked to distillation activities.