If your compiler correctly enforces security properties (and many languages do attempt to do this, at least in well-defined "safe" subsets), then of course any source code compiled with it is going to be safe. It's the difference between the JavaScript/WASM JIT-powered sandbox on the one hand, and the ActiveX code-signing-based model on the other. Which is safer?
A compiler can tell if some program will halt, for many reasonable cases. For example, if the program comes with a proof that places it in some well-defined computational complexity class, a compiler will be able to tell that it halts. And even if a program cannot be said to halt in general, it can still be said to halt conditional on some "unsafe" assertions being true. Similarly, we can verify safety properties for many useful programs, and we can even use "unsafe" assertions to limit the extent of uncertainty about these properties that we cannot directly verify.
