Anthropic says it has sharply blunted a technique rivals use to copy the abilities of its Claude AI models, announcing Monday that new safeguards have reduced data harvested in distillation attacks by about 90 to 99 per cent.
Distillation involves querying a stronger model at scale and using its answers to train a cheaper imitation. Anthropic said the reasoning data attackers can now extract is reported at 10 to 100 times lower than before, after it restricted access to Claude's protected reasoning and made probing through proxy networks harder. The company said it has banned tens of thousands of accounts linked to distillation campaigns.
Anthropic pointed to Chinese startup Z.ai, the developer of the GLM models, as one example of the activity it is targeting. Z.ai did not immediately respond to a request for comment. Anthropic said its detection carries a false-positive rate below 0.1 per cent.
The company framed distillation as both a safety and a national-security concern, arguing that copied models may strip out the guardrails built into the originals. OpenAI and Google have raised similar concerns as U.S.–China competition over advanced artificial intelligence deepens.
