• jrandomhacker@beehaw.org
    link
    fedilink
    arrow-up
    12
    ·
    21 hours ago

    LLMs fundamentally don’t/can’t have “sanitized” or “unsanitized” content - it’s all just tokens in the end. “Prompt Injection” is even a bit too generous of a term, I think.

    • businessfish@lemmy.blahaj.zone
      link
      fedilink
      English
      arrow-up
      2
      ·
      21 hours ago

      sure but one would hope that if the agent is interpreting content from the web as instructions that there would be literally any security measure between the webpage and the agent - whether that’s some input sanitization, explicit user confirmation, or prohibiting the agent from interpreting web pages as instructions at all.