Federal investigators bypassed safeguards on four popular AI chatbots, producing viable attack plans within hours.