They won't fix subtle bugs by themselves, for sure. But I have found the agentic AI to be a great assist during device reverse engineering sessions. It's fun to work on an open source project with many bits marked "unknown", and then have the AI figure out the meaning based on packet captures from known various device states. It does all the data analysis, finds the bit patterns, etc. And, yes, writes the code to use the findings. It still requires a very significant amount of human testing, and understanding of the code. But this is a task I just would not have undertaken by myself without the tool. I also don't need to know the exact intricacies of the project's language, like being forced to check for != 0 value in Go, which is not needed in C. It is readable, though, just not writable by me, because I didn't spend enough time writing Go by hand.
TLDR, the LLMs are greatly helpful tools.