If someone uses an LLM to produce the code, I'd guess they'll use it to evaluate the code as well.
The old Coverity used to achieve similar results in a different way, spotting probable mistakes based on patterns its heuristics found in the rest of the same codebase.