That's an extremely difficult task for any but the most trivial problems. It shares many of the same issues with formal verification, which has also failed to catch on beyond the odd niche project.
Differentiable Neural Computers [1] are getting us closer (by incorporating an explicit I/O operations on a memory) but it's uncertain how much test data you would need to build complex code routines.
Turning any kind of genetic / ML system loose on that? How would you even write test cases, short of backfeeding production data? And expecting the result to pass muster with auditors?
Genetic Programming is not getting much attention because it is inherently inefficient, and there are much better optimization methods out there. Genetic Programming, similarly to Monte Carlo methods, is often the "easy way out" when you are trying to solve an optimization task where you cannot (bother to) model the process.
For example: You could train Neural Networks' weights with Genetic Programming, but that's extremely inefficient compared to gradient descent.
Bottom-line: Genetic Programming is not a magic bullet, and is not guaranteed to converge to a workable solution in a "reasonable" time.
GP doesn't seem to get much attention so we are far from this method being practical. Actually, I'm puzzled why GP is so unpopular compared to say, type system research, since it could have a vastly greater impact if it were to make progress.