But it does eliminate whole classes of cheats like ESPs and wallhacks. It also allows for servers to trust the clients. There's no issue with letting a client determine if they've killed someone instead of needing on the server to rollback and check for kills.

You can still wallhack based on directional sound, and or minimap.

Image recognition is way too slow to help with aiming.

Maybe check out some existing ones before confidently making such a claim? E.g. https://github.com/petercunha/Pine. IIRC there are even high quality image recognition-based aimbots for sale; can’t find them right now.

Used to be. Moore's law isn't quite dead yet, nor are algorithmic computer vision improvements.

And when we get to the point that those are available to anyone with $40 and a credit card they'll be handled. Until then they're a theoretical attack for all intents and purposes.

This has been a thing since day 1 of Overwatch at the very least. Most aimbots being used didn't touch the executable or memory at all, they were called 'color bots' because all they had to do was match pixel color values. They were basic enough a concept you could build them in AutoIT macro software along with ramp up and down features so you weren't just instantly spamming headshots, but instead incrememently your aim would get better. They became so sophisticated that you could be sat behind someone physically using one and not be sure they were cheating or not.