A new Penn State study shows you don’t need hacking skills or prompt-engineering tricks to break past AI safety rules.