Anthropic introduces new "anti-plagiarism" measures, says it has reduced distillation attack data by over 90%.

Anthropic said Monday that new measures designed to prevent competitors from copying the capabilities of its Claude AI model 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, driven by intensifying AI competition, with companies including Anthropic, OpenAI and Alphabet Inc. Class C (GOOGL.US) all believing the practice allows rivals to replicate advanced capabilities faster and at lower cost. Anthropic said safeguards restricting access to Claude's reasoning capabilities have significantly reduced extraction attempts, with "the number of reasoning tokens extracted by distillers" now "10 to 100 times lower than before." The California-based company said it has made it harder for others to use proxy networks to repeatedly probe Claude's protected reasoning capabilities, thereby blocking extraction attempts and hindering large-scale distillation campaigns. Anthropic said its security systems are "roughly as robust as those that prevent Claude from assisting in the creation of bioweapons," and have a "false positive rate of less than 0.1%." Anthropic had previously also banned tens of thousands of accounts linked to distillation activity.