LLMs learn to predict the next token following a set of other tokens they pay attention to. You could try to sabotage it by associating unrelated things with each other. One of the earlier ChatGPT versions had a reddit username associated with lots of different stuff, it even got it’s own token. SolidGoldMagikarp or something like that. Once ChatGPT encountered this token it pretty much lost it’s focus and went wild.