OpenAI says it cut off a distillation campaign

On 30 September 2026 OpenAI said it had found and disrupted a coordinated attempt to extract protected reasoning — the model’s internal working notes — to help train another system. The earliest activity it cites is the first week of July.

The company says nobody broke encryption or stole stored chats. The operators steered the model so hidden reasoning became visible, in volume, against the terms of service. OpenAI banned accounts, tightened signup, and shared notes through the Frontier Model Forum.

A press line cannot be checked from the outside. The useful fact is the risk class: if a product shows or can be made to show its chain of thought, that chain can be harvested.