A team of eight researchers reports that encrypted reasoning blocks returned by major AI model APIs can sometimes be replayed to a compatible, less guarded model, which then transcribes the ...