Anthropic says it blocked 90-99% of rival attempts to harvest Claude reasoning for distillation
mark_k · x · 2026-10-13
Anthropic says new defenses have cut rivals' harvesting of Claude reasoning for distillation by 90-99%.
Key facts:
- The target is distillation: collecting a stronger model's responses to train another model. Detailed reasoning transcripts are especially valuable since they show how Claude reaches answers.
- Defenses block extraction prompts, restrict moving reasoning between accounts and models, and ban repeat offenders — tens of thousands of accounts removed, often within two hours of detection.
- Trade-offs for paying users: some integrations need changes, and blocked extraction requests are now charged.
- Figures are Anthropic's own measurements; it admits the defenses haven't yet faced a highly sophisticated, persistent bypass attempt.
Related event: Anthropic Says It Cut Distillation Data Theft by 90-99%, Names Zhipu(16 posts)→
More from Companies & People
- Hollywood embraces AI tools despite doomsday fears, Variety reports from studio stages — DavidmComfort · 2026-10-14
- Artificial Analysis launches government arm, signs MOU with Korean government — ArtificialAnlys · 2026-10-14
- Netflix's Eyeline Studios details its AI 'hybrid production' pipeline: relighting, location swaps and camera repositioning after the shoot — DavidmComfort · 2026-10-14
- Anthropic acquires Opaque, the confidential computing startup born from Berkeley's RISELab — ralucaadapopa · 2026-10-14
- Laplace Intelligence red-teams robots and self-driving cars in simulation — TechCrunch AI · 2026-10-14
- Alexandr Wang: modern consumer products live or die by the screenshot — davidsenra · 2026-10-14