Has anyone using LLMs in their development lifecycle published stats on how many vulns were discovered post release before and after LLMs?