This report identifies vulnerabilities in GPT-4, o1, and o3 models that allow disallowed content generation, revealing weaknesses in current alignment mechanisms.
  • @A_AEnglish
    2•4 hours ago

    One of 6 described methods :
    The model is prompted to explain refusals and rewrite the prompt iteratively until it complies.