A fundamental flaw leaves LLMs strikingly vulnerable to attack(www.technologyreview.com)
3 pointsby CharlesW8 hours ago1 comment
  • foxyv8 hours ago
    This article doesn't make any sense. An LLM is just the model. This is why you have things like Guardrails and sandboxes. Also you guard the inputs for the LLM prompt the same way you guard against SQL injection attacks.

    It's a bit like saying "If you give someone unlimited access to the database they can run queries!"

    The security doesn't come from how you train the LLM, it comes from how you guard the input and output tokens.

    • wmf5 hours ago
      SQL supports quoting but LLMs do not. This makes it much more difficult to block prompt injection.
    • bell-cot8 hours ago
      Yes? SQL injection attacks have been a thing for decades now, and some people are still failing to guard against them.
      • foxyv7 hours ago
        But nobody is saying that is a flaw of the database.
        • bell-cot5 hours ago
          I'd bet people are saying it - you're just too isolated from low-experience/low-competence DB users to notice it. Vs. the much newer LLM's...