If you want, you can think of UB like a mathematical singularity in some physical theory. The theory of the language has nothing to say about what happens in such a program. But that doesn't mean that we can't reasonably talk about what happens not using that theory [1]. Indeed, one of the reasons that UB are of concern is that some of them are frequent causes of security exploits - and that's the thing we ultimately care about, not that the program loses its semantics. But again, not all of them are equally common causes of such outcomes, and not all of them are equally hard to avoid in the first place.
[1]: In fact, this is easy to explain in software: the programming language can say nothing about the meaning of a program with UB - indeed, it has no meaning in that language - but because we do have an executable, the program still has a well-defined meaning in the machine language it compiled to (machine language has no UB, or, even if some machine architecture does declare that some instruction stream is UB, most programs with UB in some programming language still do not compile to a program with UB in machine code). So the program that has no meaning in C++ still has meaning in machine code, and as that is the program we ultimately run and care about, we can talk about which UBs are more or less likely to result in which machine code behaviour.
