Maybe a "good" solution is to make the LLM aware that it's breaking something and report back, so the exploits, 0days and everything found in the middle of the process can be fixed.