Tensor Trust is an online game that allows players to exploit prompt injection vulnerabilities in ChatGPT against other player’s preset defense instructions of user input context and valuation, for research purposes

  • Szymon@lemmy.ca
    link
    fedilink
    English
    arrow-up
    9
    arrow-down
    1
    ·
    edit-2
    1 year ago

    Help researchers develop more secure AI.

    Or are we helping researchers develop more sophisticated AI attacks?

    Edit: this is one of those things we should have regulations for… to have someone ask, “why exactly are you doing this?” and act appropriately

    • ubermeisters@lemmy.world
      link
      fedilink
      English
      arrow-up
      15
      ·
      edit-2
      1 year ago

      There’s no real difference between helping a company develop defense against attacks, and helping them develop new attacks.

      • Szymon@lemmy.ca
        link
        fedilink
        English
        arrow-up
        3
        arrow-down
        1
        ·
        1 year ago

        Fair enough, and at the end of the day it’s suspect to disguise AI training as a video game for the public. Pay me to do your work for you, or fully disclose your financers and intentions if it’s supposed to be for the greater good somehow.

    • Grimy@lemmy.world
      link
      fedilink
      English
      arrow-up
      6
      ·
      1 year ago

      We have hacker style events all the time and many websites exist that gamify it. White hat hacking for fun is a complete legitimate thing and should absolutely not be regulated.

    • jacksilver@lemmy.world
      link
      fedilink
      English
      arrow-up
      2
      ·
      1 year ago

      This also doesn’t help develop much of anything. Seems like a silly game and that’s about it.