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

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

  • Cornpop@lemmy.worldEnglish
    6·
    9 hours ago

    This is so stupid. You shouldn’t have to “jailbreak” these systems. The information is already out there with a google search.

  • meyotch@slrpnk.netEnglish
    71·
    11 hours ago

    My own research has made a similar finding. When I am taking the piss and being a random jerk to a chatbot, the bot much more frequently violates their own terms of service. Introducing non-sequitur topics after a few rounds really seems to ‘confuse’ them.