Most AI chatbots can be easily jailbroken and prompted to generate dangerous information, according to new research.