Open Menu
AllLocalCommunitiesAbout
lotide
AllLocalCommunitiesAbout
Login

ChatGPT-4o Guardrail Jailbreak: Hex Encoding for Writing CVE Exploits.

⁨80⁩ ⁨likes⁩

Submitted ⁨⁨6⁩ ⁨months⁩ ago⁩ by ⁨Dot@feddit.org⁩ to ⁨technology@lemmy.world⁩

https://0din.ai/blog/chatgpt-4o-guardrail-jailbreak-hex-encoding-for-writing-cve-exploits

source

Comments

Sort:hotnewtop
  • BaroqueInMind@lemmy.one ⁨6⁩ ⁨months⁩ ago

    This is a good read. LLMs will never be true AI, so breaking the censorship is akin to fighting back against jack-booted cops who think they know what’s best for you and that you should obey, i.e. the big corporations that run these things.

    source
    • ContrarianTrail@lemm.ee ⁨6⁩ ⁨months⁩ ago

      LLMs are true AI. AI doesn’t mean what most people think it means. AI systems from sci-fi movies like HAL 9000, JARVIS, Ava, Mother, Samantha, Skynet, and GERTY are all AI, but more specifically, they are AGI (Artificial General Intelligence). AGI is always a type of AI, but AI isn’t always AGI. Even a simple chess-playing robot is an AI, but it’s a narrow intelligence - not general. It might perform as well as or better than humans at one specific task, but this ability doesn’t translate to other tasks. AI itself is a very broad category, kind of like the term ‘plants.’

      source
      • Neon@lemmy.world ⁨6⁩ ⁨months⁩ ago

        No, AI means AI

        Corporations came up with AGI so they could call their current non-AI AI

        It’s a LLM. Not an AI.

        source
        • -> View More Comments
      • BaroqueInMind@lemmy.one ⁨6⁩ ⁨months⁩ ago

        A math model predicting language replies using a matrix is not intelligent.

        AI implies either sentience or sapience constructed outside of an organ. None of which is possible with machine learning large language models, it’s just math for now.

        source
        • -> View More Comments
  • paraphrand@lemmy.world ⁨6⁩ ⁨months⁩ ago

    It really does not feel like AGI is near when all of these holes exist. Even when they are filtered for and thus patched over, the core issue is still in the model.

    source
    • muntedcrocodile@lemm.ee ⁨6⁩ ⁨months⁩ ago

      Ironically the smarter the ai gets the harder it is to censor. Also the more u sensor it the less intelligent and less truthful it becomes.

      source
      • Telorand@reddthat.com ⁨6⁩ ⁨months⁩ ago

        the less intelligent and less truthful it becomes.

        Incorrect, because of this simple fact: garbage in, garbage out. Feed it the internet, get the internet.

        source
        • -> View More Comments
    • Blue_Morpho@lemmy.world ⁨6⁩ ⁨months⁩ ago

      Agi and LLM are two different things that fall under the general umbrella term “AI”.

      That a particular LLM can’t be censored doesn’t say anything about its abilities.

      source
  • TheFriar@lemm.ee ⁨6⁩ ⁨months⁩ ago

    Why does that thumbnail bring to mind some kind of white supremacist ceremony

    source
    • vhstape@lemmy.sdf.org ⁨6⁩ ⁨months⁩ ago

      It’s the logo of “0din”, which is a Mozilla-backed bug bounty (say that five times fast) with a focus on GenAI

      source
    • muntedcrocodile@lemm.ee ⁨6⁩ ⁨months⁩ ago

      Im assuming there is an agenda to associate uncensored ai with extremism.

      source
    • theredknight@lemmy.world ⁨6⁩ ⁨months⁩ ago

      The logo for Odin has a Nordic rune in it which is popular to white supremacists because the Nazis also co-opted them as symbols. They are not by nature about supremacy, they are an old alphabet.

      source
  • tgxn@lemmy.tgxn.net ⁨6⁩ ⁨months⁩ ago

    sure, but the source of the “Python CVE exploit” already has to exist in the AI’s training dataset, there are lots of example CVE scripts online, you could probably also find it with a quick Google.

    source