Anthropic Says Anti-Distillation Defenses Cut Stolen Data by 90-99%, Names Zhipu's GLM-5.3

On October 13, Anthropic published an article on "distillation defenses," revealing that over the past few months its new defensive methods reduced the amount of data stolen from Claude via distillation attacks by roughly 90%-99%, and that it turned away substantial revenue and paid traffic totaling trillions of tokens to do so.

Confirmed

• Anthropic officially stated that the new anti-distillation methods have cut the data harvested by distillers by about 90%-99% (m1, m3, and m4 all relay the same official figure)
• To implement this defense, Anthropic declined substantial revenue and paid requests totaling trillions of tokens (m4)
• The article called out Zhipu by name: during the development of GLM-5.3, it distilled cyber-offense-and-defense capabilities from leading US models, attempted to distill Claude Fable 5, and took follow-up actions after discovering that Anthropic's cybersecurity guardrails weakened attack effectiveness (m2, m3)
• One distiller's network controlled over 20,000 fake accounts (a specific figure within the hundreds-of-thousands range: "over 20,000"), covertly extracting Claude's reasoning outputs at scale (m5)
• Design rationale of the defense: ordinary users rarely ask a model to output full hidden reasoning, so suspected distillers' extraction requests can be detected this way; the initial guardrail release was bypassed, after which distillers scaled up (m5)

Why it matters

• Distillation attacks use vast numbers of fraudulent accounts to covertly extract leading models' capabilities at industrial scale and cheaply replicate frontier models; the effectiveness of anti-distillation directly bears on frontier labs' moats
• @abhiadesai commented that this is exactly why open-source model quality is a "lagging indicator"—open-source models' quality typically lags behind the closed-source frontier models they can distill from

2026-10-13 ~ 2026-10-13 · 5 related posts

• Anthropic says new defenses cut distillation attack data harvesting by 90-99%, cites Zhipu's GLM-5.3 — AnthropicAI · 2026-10-13
• Anthropic says it cut distiller data harvesting by 90-99%, turning away trillions of tokens — abhiadesai · 2026-10-13
• Anthropic Says Zhipu Distilled Claude to Train GLM-5.3, Calls It Top Cyber Open-Weights Model — No-Point1424 · 2026-10-13
• Leaked Anthropic anti-distillation details: one network ran 20,000+ fake Claude accounts — teortaxesTex · 2026-10-13

1 near-duplicate retellings: scaling01