I’m writing a paper on this, actually. Basically, it’s okay-ish at it, but has definite blind spots. The most promising route is to have AI use a traditional static analysis tool, rather than evaluate the code directly.

That’s the kind of logic people historically used when designing low level programming languages. It’s not the kind of logic you should use or that people nowadays usually do use. Undefined behavior is widely seen as a Bad Thing in the programming language design community.