OpenAI says Moonshot AI-linked individuals led campaign making 16,000 attempts to extract hidden reasoning
kimmonismus · x · 2026-10-01
- OpenAI disclosed a large-scale campaign to extract its models' hidden reasoning, saying individuals linked to Moonshot AI (the developer of Kimi) were behind a core part of it.
- Over two days, OpenAI logged 16,000 extraction attempts from 4,000+ users; further investigation tied related activity to 15,000+ users.
- Operators tried moving encrypted reasoning between conversations and prompting the model to reveal its contents, since hidden reasoning is valuable training material for competing models.
Related event: OpenAI accuses Moonshot AI of distilling its hidden model reasoning(6 posts)→
More from Models
- Perplexity's new contextual embedding model tops turbopuffer's context-bench with 50%+ recall gain — marktenenholtz · 2026-10-01
- Local LLM Fans Face a Reality Check: 8-10T Parameter Models Will Need a Server Rack — power97992 · 2026-10-01
- Inception launches Mercury Decide, a structured decision model hitting 14 decisions per second — StefanoErmon · 2026-10-01
- OpenAI Gives Daybreak Blue Users One Day to Get a FIDO2 Key or Lose Access — Admirable_Trainer_54 · 2026-10-01
- Experiment restricts Qwen to Wikipedia's 250 most common tokens plus rare ones — cephaloform · 2026-10-01
- Strix Halo Plus R9700 Hybrid Hits 60+ tok/s, Beating DGX Spark on Local LLM Inference — darklordfireape · 2026-10-01