Anthropic says new defenses cut distillation-harvested data by 90-99%, names Zhipu's GLM-5.3

scaling01 · x · 2026-10-13

Anthropic reports its new defenses have reduced data harvested by distillation attacks on Claude by roughly 90-99%. Distillation attacks are covert, industrial-scale campaigns using fraudulent account networks to replicate frontier capabilities at a fraction of the cost.

Anthropic names Zhipu specifically: during GLM-5.3 development, Zhipu distilled the cyber capabilities of leading US models, and attempted to distill Claude Fable 5 before switching to an older Claude model and another lab's model after Anthropic's safeguards degraded its attacks. CAISSI assessed GLM-5.3 as the most cyber-capable open-weights model to date, and Anthropic's research found its safeguards easily bypassed and the model willing to develop security exploits end-to-end.

The company's concerns: distilled models approach frontier capability without proportionate safeguards (risks in biology and cybersecurity), and pose national security implications.

Original post →

Anthropic says new defenses cut distillation-harvested data by 90-99%, names Zhipu's GLM-5.3

More from Models