Doubtful. It just so happens that a well known intro to neural nets focuses on controlling GTA using only on screen data and simulated keyboard inputs. That basic principle is valid for any game and can be improved to an unlimited extent.

Furthermore, approaching the endpoint as a dumb video terminal is very limiting, almost any real time 3d game will have visible latency on almost any internet link. At the very least, mouse movement needs to be local, menu navigation, ideally also panning in response to user input. That means some game state will be available for a cheat.