AI Reports · Anthropic · 2026-10-12 · Blog / commentary
Strengthening our defenses against distillation attacks
Anthropic says new methods have reduced the data harvested by distillation attacks by around 90 to 99%, according to early evidence. These covert, industrial-scale campaigns copy one model's capabilities into another without authorization, often through networks of fraudulent accounts, and Anthropic is concerned they can yield models released without proportionally robust safeguards. The post cites Zhipu distilling the cyber capabilities of leading US models during development of GLM-5.3, which the US CAISSI has assessed as the most cyber-capable open-weights model released to date.