“As of a month prior to the interview, a prompt injection technique involving translating a harmful request to Spanish and then Base64 encoding it was successful in bypassing ChatGPT's safety filters.”