I see. I guess this does solve the issue of a program managing to insert arbitrary machine code on an address and point the CPU at it, or replacing the memory where the machine code already exists. But I think there already exists OS-level protections against writable executable memory?

It also doesn't solve one important attack vector: the existing executable has logic that can be used maliciously (e.g. a function that sends a file to the internet), and this logic is executed based on data on memory. In this case you only need to manipulate the data, without manipulating the "code", in order to exploit it.

I'm not sure I understand why would you need special hardware for that.

I think that the principle of "code is data" is that data influences the algorithms either way, so there is not much of a point in separating the two.