AI chatbots' safeguards can be easily bypassed, say UK researchers