Moravec's paradox likely comes in to play, what's easy is hard and vice versa.

The puzzles would probably be easy. Myst's puzzles are basically IQ tests, and LLMs ace traditional IQ tests: https://trackingai.org/home

On the other hand, navigating the environment, I think the models may fail spectacularly. From what we've seen from Claude Plays Pokemon, it would get in weird loops and try to interact with non-interactive elements of the environment.