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