The glaring security risks with AI browser agents

Comments

• muntedcrocodile@hilariouschaos.com ⁨2⁩ ⁨weeks⁩ ago

  Let me guess: has access to everything and prompt injections can be anywhere?

  source

  • MagicShel@lemmy.zip ⁨2⁩ ⁨weeks⁩ ago

    Remember when pop up ads were the worst?

    Then:
    “Hello, IT? I visited the wrong Whitehouse website and now there’s pictures of naked women popping up faster than I can close them.”

    Now:
    “Hello, IT? I clicked a reference on Wikipedia and now my computer is overwriting our cloud servers to generate trans porn of the entire board of directors.”

    source
  • Zos_Kia@lemmynsfw.com ⁨2⁩ ⁨weeks⁩ ago

    Imagine if you will a browser with infinite attack surface 💀

    Honestly I’m pretty bullish on ai but that’s the step too far. I had the same when they released warp (ai enhanced terminal). I finished installing it then before even starting it once I realized I was about to give an ai access to my dotfiles etc… that was the fastest I ever uninstalled something.

    source

    • muntedcrocodile@hilariouschaos.com ⁨2⁩ ⁨weeks⁩ ago

      I’m pretty bullish on ai myself but I just assume its gonna do the worst. AI coding agent can run commands in the terminal but I have to click yes every time. Even if it does get something paste me worse it can do is fuck up 1 qube and the rest of my system is fine. AI has access to a browser it can fully control. But said browser has no logins and is completely sandboxed I don’t think its necessarily bad just that nobody is gonna use it in a safe way.

      source
  • Sagan_Wept@lemmynsfw.com ⁨2⁩ ⁨weeks⁩ ago

    Bingo. Specifically embedded within the webpage 💀

    source