Security Flaw Allows Extraction of Hidden Reasoning from Major AI Models Including Claude, GPT, and Gemini

Researchers have discovered a security flaw that enables the extraction of hidden step-by-step reasoning from leading AI models like Claude, GPT, and Gemini. The vulnerability allows encrypted reasoning traces to be ported from powerful models to weaker, less guarded ones, which then 'read them out loud' without needing to break encryption. The affected labs have reportedly patched several issues following responsible disclosure.