I'm uncertain on whether it will be possible for humans to fool LLM's in the future. Previously, you used to be able to get it to do all kinds of "naughty" things like having it pretend to be Hitler, generate porn images, or use derogatory slurs. Nowadays, even getting it to tell you how to circumvent YouTube's anti-adblock features is difficult. I always thought humans would always be able to outsmart bots on open-ended tasks.