A team of eight researchers reports that encrypted reasoning blocks returned by major AI model APIs can sometimes be recovered.