The LLMs have injected enough training data that they can find something that looks rather close to a problem that's been solved once before and use that in most cases. It won't be perfect, but a human can usually tune it from there. I'm not sure if I'd trust a security audit from an LLM because they're incapable of doing actual analysis and rely on past data that's going to be heavily weighted towards saying that something is good. That may have been true in isolation, but not when the LLM has combined two pieces of code in a way that creates a vulnerability.
It will be interesting if the role of humans in software development shifts away from writing code and to verifying its correctness. There will still be a need for some human developers as no matter how good an LLM gets it can't produce output that wasn't a part of its training data. For computer games, there's a lot of repetition and getting one to spit out a 3D game engine isn't hard when it's been able to train off of thousands of examples. I'd say that building a novel experience for a what it would struggle with, but it seems that most people buying games largely want more of the same, so perhaps that's not a problem.
