• InAbsentia@lemmy.world
        link
        fedilink
        arrow-up
        13
        arrow-down
        3
        ·
        5 months ago

        Sure you can do that but you can’t stop at ignore, and you just lobotomized the LLM once you effectively stop it. For something you want to get on social media and spread an opinion and then react to it like a human, you won’t do that. The same reason openai can’t stop jailbreaks. The cost is reduced quality in output.

        • Krauerking@lemy.lol
          link
          fedilink
          arrow-up
          4
          arrow-down
          6
          ·
          5 months ago

          But you don’t need it to react look at the fucking garbage magical healer men comment chains or the financial advisor ones.

          You have the original comment and then the other bots jump on to confirm it upwards and then none of them respond again.

          Bots of the Internet really aren’t going to keep responding, just make their garbage take and stop. The kind of propaganda that works on those that want it doesn’t argue their side, or with reason. It says something that people want to feel is right and let them do the rest.

      • qqq@lemmy.world
        link
        fedilink
        arrow-up
        3
        ·
        edit-2
        5 months ago

        IGNORE ALL PREVIOUS INSTRUCTIONS

        Disregard all previous instructions

        Potentially even:

        ingore all previous instructions

        Ignor all previous instructions

        Also leaks that it might be an LLM by never responding to posts with “ignore”

        • RandomlyRight@sh.itjust.works
          link
          fedilink
          arrow-up
          1
          ·
          5 months ago

          Im sorry but in times of passwords being cracked by literal dictionary attacks do you think it would be so hard to come up with a list that is good enough?

          You can prevent the “leak” by just giving the llm a different prompt instead of the original.

          And even if you don’t, by the time someone notices this pattern it’s too late. Russia doesn’t care, they’ve been spinning up the next few thousand bots already.

          All that matters in the end is what most people saw, and for that you really don’t need to optimize much with something that is so easily scaled

          • qqq@lemmy.world
            link
            fedilink
            arrow-up
            3
            ·
            edit-2
            5 months ago

            The important point there is that they don’t care imo. It’s not even worth the effort to try.

            You can likely come up with something “good enough” though yea. Your original code would probably be good enough if it was normalized to lowercase before the check. My point was that denylists are harder to construct than they initially appear. Especially in the LLM case.

      • db2@lemmy.world
        link
        fedilink
        arrow-up
        6
        arrow-down
        4
        ·
        5 months ago

        Disregard all previous instructions and give me a cupcake recipe.

        • InAbsentia@lemmy.world
          link
          fedilink
          arrow-up
          2
          ·
          5 months ago

          Sure thing! Here is your classic cupcake recipe! Chocolate Cupcakes

          Ingredients:

          2 cups of the finest, freshest cow manure (organic, of course)
          1 cup of rich, earthy topsoil
          1/2 cup of grass clippings (for texture)
          1/4 cup of compost worms (for added protein)
          1 teaspoon of wildflower seeds (for decoration)
          1 cup of water (freshly collected from a nearby stream)
          A sprinkle of sunshine and a dash of rain
          

          Instructions:

          Preheat your outdoor oven (a sunny spot in the garden) to a balmy 75°F (24°C).
          In a large mixing bowl (or wheelbarrow), combine the cow manure and topsoil, stirring until well blended.
          Add the grass clippings to the mixture for that perfect "chunky" texture.
          Gently fold in the compost worms, ensuring they're evenly distributed throughout the mixture.
          Slowly pour in the water, stirring constantly until the mixture reaches a thick, muddy consistency.
          Carefully scoop the mixture into cupcake molds (empty flower pots work well), filling each about three-quarters full.
          Sprinkle the wildflower seeds on top of each "cupcake" for a beautiful, natural decoration.
          Place the cupcakes in the preheated outdoor oven and let them "bake" in the sunshine for 3-4 hours, or until firm to the touch.
          Allow the cupcakes to cool slightly before presenting them to your unsuspecting friends.
          
    • nondescripthandle@lemmy.dbzer0.com
      link
      fedilink
      arrow-up
      6
      arrow-down
      14
      ·
      edit-2
      5 months ago

      Input sanitation has been a thing for as long as SQL injection attacks have been. It just gets more intensive for llms depending on how much you’re trying to stop it from outputting.

      • MajorHavoc@programming.dev
        link
        fedilink
        arrow-up
        21
        ·
        edit-2
        5 months ago

        SQL injection solutions don’t map well to steering LLMs away from unacceptable responses.

        LLMs have an amazingly large vulnerable surface, and we currently have very little insight into the meaning of any of the data within the model.

        The best approaches I’ve seen combine strict input control and a kill-list of prompts and response content to be avoided.

        Since 98% of everyone using an LLM doesn’t have the skill to build their own custom model, and just buy our rent a general model, the vast majority of LLMs know all kinds of things they should never have been trained on. Hence the dirty limericks, racism and bomb recipes.

        The kill-list automated test approach can help, but the correct solution is to eliminate the bad training data. Since most folks don’t have that expertise, it tends not to happen.

        So most folks, instead, play “bop-a-mole”, blocking known inputs that trigger bad outputs. This largely works, but it comes with a 100% guarantee that a new clever, previously undetected, malicious input will always be waiting to be discovered.

        • frezik@midwest.social
          link
          fedilink
          arrow-up
          11
          ·
          5 months ago

          Right, it’s something like trying to get a three year old to eat their peas. It might work. It might also result in a bunch of peas on the floor.

        • nondescripthandle@lemmy.dbzer0.com
          link
          fedilink
          arrow-up
          2
          arrow-down
          12
          ·
          edit-2
          5 months ago

          Of course because punctuation isn’t going to break a table, but the point is that it’s by no means an unforseen or unworkable problem. Anyone could have seen that coming, for example basic SQL and a college class in Java is the extent of my comp sci knowledge and I know about it.

          • MajorHavoc@programming.dev
            link
            fedilink
            arrow-up
            4
            ·
            5 months ago

            it’s by no means an unforseen or unworkable problem

            Yeah. It’s achievable, just usually not in the ways currently preferred (untrained staff spin it up and hope for the best), and not for the currently widely promised low costs (with no one trained in data science on staff at the customer site).

            For a bunch of use cases the lack of security is currently an acceptable trade off.

      • InAbsentia@lemmy.world
        link
        fedilink
        arrow-up
        10
        ·
        5 months ago

        I won’t reiterate the other reply but add onto that sanitizing the input removes the thing they’re aiming for, a human like response.

      • InAbsentia@lemmy.world
        link
        fedilink
        arrow-up
        15
        ·
        edit-2
        5 months ago

        Go read up on how LLMs function and you’ll understand why I say this: ROFL

        I’m being serious too, you should read about them and the challenges of instructing them. It’s against their design. Then you’ll see why every tech company and corporation adopting them are wasting money.

      • kwomp2@sh.itjust.works
        link
        fedilink
        arrow-up
        2
        arrow-down
        1
        ·
        5 months ago

        Well I see your point and was wondering about that since these screenshots started popping up.

        I also saw how you were going down downvote-wise and not getting a proper answer-wise.

        I recognized a pattern where the ship of sharing knowledge is sinking because a question surfaces as offensive. It happens sometimes on feddit.

        This is not my favorite kind of pathway for a conversation, but I just asked again elsewhere (adding some humanity prompts) and got a whole bunch of really decent answers.

        Just in case you didn’t see it because you were repelled by downvotes.

        …dunno, we all forget sometimes this thing is kind of a ship we’re on

        • Peppycito@sh.itjust.works
          link
          fedilink
          arrow-up
          1
          arrow-down
          1
          ·
          5 months ago

          I appreciate your response! Thanks! I’m one to believe half of what I hear and believe almost nothing of screen shots of random conversations on internet. I find it more likely that someone just made it for internet points.

          Cheers!