If memory safety issues are 75% of exploited zero days it sounds to me like they're the biggest issue in the ecosystem by far.
Most exploited code probably exists in the application layer in a high-level, memory safe language. I would wager that but I don’t have time to cite ten papers on HN.
It's a bit like saying you should program in C, because formal verification tool X generates C code hence C is safe.
I think formal verification is the way to go with AI moving forward.