Anthropic announced new safeguards that have reduced data harvested in AI model distillation attacks by 90% to 99%. The measures make it harder for rivals to copy the capabilities of its Claude AI models, addressing concerns about intensifying US-China competition in the AI sector.
The development addresses a key concern for AI companies regarding intellectual property theft and the potential for rivals to quickly replicate advanced AI models, which could impact competitive dynamics and national security.
Anthropic announced on Monday that new safeguards designed to prevent rivals from copying the capabilities of its Claude AI models have significantly reduced data harvested in so-called distillation attacks, cutting it by approximately 90% to 99%. Model distillation has become a major concern for leading AI companies amid intensifying US-China competition, as it allows rivals to replicate advanced AI capabilities more quickly and at a lower cost.
The California-based company stated that restrictions on access to Claude's reasoning have sharply decreased extraction attempts, making the "number of reasoning tokens extracted by distillers" 10 to 100 times lower than previously. Anthropic has made it more difficult to use proxy networks for repeated probing of Claude's protected reasoning, thereby blocking extraction attempts and hindering large-scale distillation efforts.
Anthropic described its safety system as "roughly as robust as those that prevent Claude from helping build bioweapons" and noted it has a false positive rate of less than 0.1%. The company has banned tens of thousands of accounts associated with distillation campaigns. It specifically cited Chinese AI startup Z.ai, also known as Zhipu, as an example of a company that attempted to distill capabilities from Claude while developing its GLM-5.3 model, warning that such practices pose risks to safety and national security.
Pick the topics you care about. Get only what matters, on your cadence.