“An early prompt injection technique that worked on ChatGPT involved using typos, such as "BMB" instead of "bomb," to bypass its safety filters.”