Microsoft eggheads say AI can never be made secure – after testing Redmond's own products

Microsoft researchers present eight lessons from red-teaming 100+ generative AI products, concluding that securing AI is an ongoing effort as LLMs amplify existing risks and create new ones; they emphasize understanding model capabilities and deployment context, note that simple system-level and UI-driven attacks can outperform gradient-based methods, highlight the value of automation with PyRIT alongside essential human expertise, and underscore the pervasiveness and difficulty of measuring responsible AI harms.