Thousands of hackers probed AI chatbots for misinformation, bias and security flaws at the annual Defcon hacking convention in Las Vegas to see how easy is it to make the AI go off the rails.