It seems clear that LLMs significantly change threat model math, but this observation alone does not explain how or why; the asymmetry that you’re describing is a property of pre-LLM software as well.