Definition
A newly disclosed attack that lets outsiders decrypt and read an AI model's private 'internal reasoning' — the step-by-step notes it normally hides from users while working out an answer. It affects the three largest AI providers at once, showing the weakness is in a shared design pattern, not one company's mistake.
Why it matters
Companies rely on hidden reasoning staying hidden both for competitive/IP reasons and for safety filtering; if attackers can read it, they can extract proprietary logic and potentially find ways around safety checks that were meant to happen 'invisibly.'