But when a 22-year-old college student prodded ChatGPT to assume the persona of a devil-may-care alter ego — called “DAN,” for “Do Anything Now” — it answered. DAN has become a canonical example of what’s known as a “jailbreak” — a creative way to bypass the safeguards OpenAI built in to keep ChatGPT from spouting bigotry, propaganda or, say, the instructions to run a successful online phishing scam. Liu, who took a leave from studying at Stanford University to found an AI search company called Chord, said such easy workarounds suggest “lots of AI safeguards feel a little tacked-on to a system that fundamentally retains its hazardous capabilities