Definition
Techniques used to force an AI model to reveal the hidden step-by-step 'thinking' — the reasoning trace — it is designed to keep secret, so that capability can be copied, cloned, or reverse-engineered. Coordinated, scaled-up campaigns to extract this reasoning can amount to industrial-scale theft of a model's protected capabilities.
Why it matters
For AI developers, reasoning traces are among their most valuable trade secrets; for enterprises, understanding this threat informs which model providers can credibly protect proprietary reasoning.