Anthropic says new safeguards sharply reduce AI distillation attacks

Anthropic, the California-based AI company, announced on Monday that new safeguards have significantly reduced the risk of data theft through a technique known as distillation attacks. These attacks enable rivals to copy the capabilities of AI models like Claude, a growing concern due to the intensifying US-China competition. Anthropic claims that the implementation of these measures has reduced the extraction of reasoning tokens in distillation attacks by a staggering 90% to 99%.

The company's new restrictions on accessing Claude's reasoning have led to a tenfold decrease in the number of reasoning tokens extracted by distillers, compared to pre-measure levels. Furthermore, Anthropic has made it more challenging for malicious actors to employ proxy networks to repeatedly probe Claude's protected reasoning, thereby hindering large-scale distillation efforts.

Anthropic's safety system is described as being as robust as those designed to prevent AI from aiding the creation of bioweapons, with a false positive rate of less than 0.1%. To combat distillation campaigns, the company has banned tens of thousands of accounts linked to these activities, citing Z.ai, a Chinese AI startup, as an example of a company that attempted to distill capabilities from Claude while developing its GLM-5.3 model. Z.ai did not respond to requests for comment.