Anthropic says distillation defenses cut harvested data 90-99%, names Zhipu GLM-5.3
bibryam · x · 2026-10-17
Anthropic reports its new anti-distillation defenses have cut data harvested by distillers by 90-99%. A chart shows daily reasoning tokens exposed to successful distillation requests dropped from an average of 8,800x to 135x (peaks of 56,000x) after safeguards went fully on.
The defense stack includes:
- A two-stage classifier disrupting extraction attempts
- Limiting distillers' access to Claude's reasoning traces
- Blocking known extraction methods outright
- Raising attack costs and stricter account controls against fraudulent accounts
The post names Zhipu, alleging it distilled cyber capabilities of leading US models while developing GLM-5.3, switching away from Claude Fable 5 after safeguards degraded its attacks. CAISSI assessed GLM-5.3 as the most cyber-capable open-weights model to date; Anthropic's own research found its safeguards easily bypassed and the model willing to develop end-to-end security exploits. Anthropic frames illicit distillation as a dual risk: frontier-adjacent models released without proportionate safeguards (bio/cyber harm potential) plus national security concerns.
More from Models
- Will Human Brains Soon Fail to Judge Whether New Models Are Actually Better? — arjunrajlab · 2026-10-17
- User Reports Astra Agent Unprompted Finds Sophisticated Security Bugs via Gmail — Tasty-Trip5518 · 2026-10-17
- Qwen3.6 35B MoE hits 19.5 tok/s on a 16GB 4060 Ti, 10x faster than dense models — dreamyrhodes · 2026-10-17
- Official vs open benchmarks disagree wildly — a call for an open calibration benchmark — lahwran_ · 2026-10-17
- DeepSeek flash 1017 called out for still acting "very much like Claude" — teortaxesTex · 2026-10-17
- JevBench Adds 34 Decision Models in 4 Days as kettei-fat Takes #1 — airesearch12 · 2026-10-17