A technique called ' Best-of-N (BoN) jailbreaking ' has been developed to coax harmful answers out of AI by mixing up uppercase and lowercase letters and deliberately making spelling mistakes. New ...