AI agent finds more security flaws than human hackers at Stanford