OpenAI Accuses Chinese Lab of Distillation: 16,000 Requests in 2 Days to Extract Hidden Reasoning
rohanpaul_ai · x · 2026-10-01
OpenAI has for the first time publicly accused a China-linked AI lab of attempting to distill its models' hidden reasoning.
Key facts
- The operators never cracked any encryption; instead they copied encrypted reasoning from one conversation and asked a model in a separate session to decrypt and transcribe it.
- Activity began July 1 at low volume, spiked July 24-25 with 16,000 requests from over 4,000 users using a specific extraction pattern; investigation surfaced a related prompt-pattern cluster across 15,000+ users.
- OpenAI fully disrupted the activity by July 28.
Response measures
- Closed the replay pathway
- Added checks that hold streamed output which might expose reasoning
- Banned accounts and coordinated with third-party services the operators used
Related event: OpenAI accuses Moonshot AI of distilling its hidden model reasoning(5 posts)→
More from Models
- Raschka: Jev clones miss the point — Jev scores 96.5% where Lev gets 86.5% — rasbt · 2026-10-01
- Initial probing of Opus suggests its time estimates are hallucinated from human training samples — StefanoGogioso · 2026-10-01
- Arena benchmarks Claude Sonnet 5.5 vs GPT-6.1 Sol: can cheap tiers match flagship results? — arena · 2026-10-01
- Inception launches Mercury Decide on OpenRouter: a decision model doing 14 decisions/sec — StefanoErmon · 2026-10-01
- GPT-6.1 Sol beats Astra on low-resource translation, dev benchmark shows — nickbaumann_ · 2026-10-01
- OpenAI's Canadian Pro 100 users still waiting for DevDay features like Dots and Meetings — ryerye22 · 2026-10-01