Research Shows AI Still Needs Human Guidance to Discover New Hacking Methods

Web security researcher James Kettle presented research at the Black Hat conference examining AI's capabilities in developing new attack methods autonomously. Kettle concluded that while AI has limited capabilities to invent entirely new attack paths without guidance, when combined with human direction at key moments, it becomes an extremely powerful tool for uncovering innovative hacking strategies. During months of experiments with models from Anthropic and OpenAI, Kettle identified an entirely new vulnerability area called "Shared-Parser Confusion," which exploits how web servers use shared code to process both requests and responses, representing a significant attack surface.