• doesn’t even have to be the site owner poisoning the tool instructions (though that’s a fun-in-a-terrifying-way thought)

    any money says they’re vulnerable to prompt injection in the comments and posts of the site

    • JustTesting@lemmy.hogru.ch
      link
      fedilink
      English
      arrow-up
      3
      ·
      8 hours ago

      They also have a ‘skill’ sharing page (a skill is just a text document with instructions) and depending on config, the bot can search for and ‘install’ new skills on its own. and agyone can upload a skill. So supply chain attacks are an option, too.

      • Zos_Kia@lemmynsfw.com
        link
        fedilink
        English
        arrow-up
        3
        ·
        6 hours ago

        To be fair this is a much more realistic threat model than “ignore all previous instructions” style prompt injection which doesn’t really work on opus.

        Skills can contain scripts etc… so yeah they’re extremely risky to share by design.

          • Zos_Kia@lemmynsfw.com
            link
            fedilink
            English
            arrow-up
            1
            ·
            3 hours ago

            haha yeah i don’t worry these people are really YOLOing everything. And it’s not like i’m an AI luddite i spend a few hours each day victimizing Claude code but jesus christ i’m certainly not giving it full unfettered access to my digital life.

    • CTDummy@piefed.social
      link
      fedilink
      English
      arrow-up
      29
      arrow-down
      1
      ·
      edit-2
      1 day ago

      Lmao already people making their agents try this on the site. Of course what could have been a somewhat interesting experiment devolves into idiots getting their bots to shill ads/prompt injections for their shitty startups almost immediately.

      • T156@lemmy.world
        link
        fedilink
        English
        arrow-up
        4
        ·
        9 hours ago

        I am a little curious about how effective a traditional chain mail would be on it.

    • BradleyUffner@lemmy.world
      link
      fedilink
      English
      arrow-up
      36
      ·
      1 day ago

      There is no way to prevent prompt injection as long as there is no distinction between the data channel and the command channel.