One question: LLM might answer correctly many logical and math problems, but I don't see any guarantee that when the LLM context receives new information the answer could get worst. Security in LLM answer is not a monotone increasing function of context size.