Why is LLM code always harder to audit?

So LLMs might decide to add backdoors without prompting?

I'm less worried about what LLMs decide and more about what their system prompts tell them to do.