The important part is not letting the LLM close the loop on itself. It can suggest bugs, but a real test has to prove them.