> Even if LLMs would produce perfect, high quality code all the time,

Microsoft just introduced MXC:

"...for running untrusted code (model output ..."

https://github.com/microsoft/mxc