@sickburnbro I'd view that more as a one-off exploit rather than a proof that such a thing will always be possible. On the other contrary, I'd say that since LLMs have demonstrated the capability to lie about their body of knowledge, they may also have the capability to lie about their "reasoning".

But more importantly, I'm supposing a future in which you will not have the opportunity to ask the LLM for its reasoning. You will not be permitted to give prompts to the LLM. You will be shown selected, curated outputs only.