Do any of these bots use their own previous outputs as further training data? That's one way these exploits could spread beyond "the same user who asks the bot to do ...