Comment by imustachyou

I’m missing something, won’t the input to the llm necessarily be plaintext? And the output too? Then, as long as the llm has logs, the real input by users will be available somewhere in their servers