The only realistic solution to this problem is to move the audio and video rendering server-side and stream final output to clients from the edge.
You have to attack this from an information theory perspective. If the client receives the information in any manner whatsoever, you must assume it is being used in the most adversarial way possible.
The AI problem is more complex, but that's more a function of how deep the gameplay is than how good the hacker is. Time constraints make pure machine vision bots subpar for fast-paced games that also have a complex tactical element.