The researchers are working with a technique identified as adversarial instruction to stop ChatGPT from allowing people trick it into behaving terribly (generally known as jailbreaking). This work pits a number of chatbots versus one another: just one chatbot plays