Interesting fact from this article: 12% of vulnerabilities in human-written code are high or critical severity. For LLM code slop, that number is 32% and has apparently been pretty constant for a while rather than improving