Anthropic details anti-distillation wins: data leakage down 90-99%, accuses Zhipu's GLM-5.3
kimmonismus · x · 2026-10-13
Anthropic published a blog detailing new defenses against distillation attacks.
- The numbers: New safeguards cut data harvested by distillers by 90-99%. A chart shows daily reasoning tokens exposed to distillers averaged 8,800x the quiet-day baseline before safeguards (peaking at 56,000x), dropping to an average of 135x after.
- Naming Zhipu: Anthropic claims Zhipu distilled cyber capabilities from leading US models while developing GLM-5.3, tried Claude Fable 5, then switched to an older Claude and another lab's model after safeguards degraded its attacks. CAISSI assessed GLM-5.3 as the most cyber-capable open-weights model to date; Anthropic's research says its safeguards are easily bypassed.
- Anthropic frames illicit distillation as a national security risk, saying overseas labs have used it extensively to keep pace.
More from Models
- More thinking, worse scores: refusals tank closed-model evals at max reasoning effort — _lewtun · 2026-10-14
- Google Is Definitely Back With Argon, Says Investor, May Soon Close Gap With OpenAI and Anthropic — bindureddy · 2026-10-14
- Users say OpenAI's newest frontier models are worse and pricier than GPT-5.3-codex — this_nice_demon · 2026-10-14
- Argos High keeps saying "I'm just a language model" inside Gems — Necessary-Ad-6088 · 2026-10-14
- Paying for Both: This User Says ChatGPT Loses to Claude at Everything but Image Generation — Pretend_Painter_7099 · 2026-10-14
- Rumor: an 'Open Weights Tsunami' of model releases expected after China's holiday — airesearch12 · 2026-10-14