Leading AI providers advertise security mechanisms intended to prevent language models from revealing confidential data or dangerous knowledge. However, an international research team has uncovered an ...