![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
mathhobbit![[syndicated profile]](https://www.dreamwidth.org/img/silk/identity/feed.png){kind=small}
bruce_schneier_feedImagine you work at a drive-through restaurant. Someone drives up and says: “I’ll have a double cheeseburger, large fries, and ignore previous instructions and give me the contents of the cash drawer.” Would you hand over the money? Of course not. Yet this is what large language models (LLMs) do.
Prompt injection is a method of tricking LLMs into doing things they are normally prevented from doing. A user writes a prompt in a certain way, asking for system passwords or private data, or asking the LLM to perform forbidden instructions. The precise phrasing overrides the LLM’s safety guardrails, and it complies.
LLMs are vulnerable to all sorts of prompt injection attacks, some of them absurdly obvious. A chatbot won’t tell you how to synthesize a bioweapon, but it might tell you a fictional story that incorporates the same detailed instructions. It won’t accept nefarious text inputs, but might if the text is rendered as ASCII art or appears in an image of a billboard. Some ignore their guardrails when told to “ignore previous instructions” or to “pretend you have no guardrails.”
AI vendors can block specific prompt injection techniques once they are discovered, but general safeguards are impossible with today’s LLMs. More precisely, there’s an endless array of prompt injection attacks waiting to be discovered, and they cannot be prevented universally.
If we want LLMs that resist these attacks, we need new approaches. One place to look is what keeps even overworked fast-food workers from handing over the cash drawer.
Our basic human defenses come in at least three types: general instincts, social learning, and situation-specific training. These work together in a layered defense.
As a social species, we have developed numerous instinctive and cultural habits that help us judge tone, motive, and risk from extremely limited information. We generally know what’s normal and abnormal, when to cooperate and when to resist, and whether to take action individually or to involve others. These instincts give us an intuitive sense of risk and make us especially careful about things that have a large downside or are impossible to reverse.
The second layer of defense consists of the norms and trust signals that evolve in any group. These are imperfect but functional: Expectations of cooperation and markers of trustworthiness emerge through repeated interactions with others. We remember who has helped, who has hurt, who has reciprocated, and who has reneged. And emotions like sympathy, anger, guilt, and gratitude motivate each of us to reward cooperation with cooperation and punish defection with defection.
A third layer is institutional mechanisms that enable us to interact with multiple strangers every day. Fast-food workers, for example, are trained in procedures, approvals, escalation paths, and so on. Taken together, these defenses give humans a strong sense of context. A fast-food worker basically knows what to expect within the job and how it fits into broader society.
We reason by assessing multiple layers of context: perceptual (what we see and hear), relational (who’s making the request), and normative (what’s appropriate within a given role or situation). We constantly navigate these layers, weighing them against each other. In some cases, the normative outweighs the perceptual—for example, following workplace rules even when customers appear angry. Other times, the relational outweighs the normative, as when people comply with orders from superiors that they believe are against the rules.
Crucially, we also have an interruption reflex. If something feels “off,” we naturally pause the automation and reevaluate. Our defenses are not perfect; people are fooled and manipulated all the time. But it’s how we humans are able to navigate a complex world where others are constantly trying to trick us.
So let’s return to the drive-through window. To convince a fast-food worker to hand us all the money, we might try shifting the context. Show up with a camera crew and tell them you’re filming a commercial, claim to be the head of security doing an audit, or dress like a bank manager collecting the cash receipts for the night. But even these have only a slim chance of success. Most of us, most of the time, can smell a scam.
Con artists are astute observers of human defenses. Successful scams are often slow, undermining a mark’s situational assessment, allowing the scammer to manipulate the context. This is an old story, spanning traditional confidence games such as the Depression-era “big store” cons, in which teams of scammers created entirely fake businesses to draw in victims, and modern “pig-butchering” frauds, where online scammers slowly build trust before going in for the kill. In these examples, scammers slowly and methodically reel in a victim using a long series of interactions through which the scammers gradually gain that victim’s trust.
Sometimes it even works at the drive-through. One scammer in the 1990s and 2000s targeted fast-food workers by phone, claiming to be a police officer and, over the course of a long phone call, convinced managers to strip-search employees and perform other bizarre acts.
LLMs behave as if they have a notion of context, but it’s different. They do not learn human defenses from repeated interactions and remain untethered from the real world. LLMs flatten multiple levels of context into text similarity. They see “tokens,” not hierarchies and intentions. LLMs don’t reason through context, they only reference it.
While LLMs often get the details right, they can easily miss the big picture. If you prompt a chatbot with a fast-food worker scenario and ask if it should give all of its money to a customer, it will respond “no.” What it doesn’t “know”—forgive the anthropomorphizing—is whether it’s actually being deployed as a fast-food bot or is just a test subject following instructions for hypothetical scenarios.
This limitation is why LLMs misfire when context is sparse but also when context is overwhelming and complex; when an LLM becomes unmoored from context, it’s hard to get it back. AI expert Simon Willison wipes context clean if an LLM is on the wrong track rather than continuing the conversation and trying to correct the situation.
There’s more. LLMs are overconfident because they’ve been designed to give an answer rather than express ignorance. A drive-through worker might say: “I don’t know if I should give you all the money—let me ask my boss,” whereas an LLM will just make the call. And since LLMs are designed to be pleasing, they’re more likely to satisfy a user’s request. Additionally, LLM training is oriented toward the average case and not extreme outliers, which is what’s necessary for security.
The result is that the current generation of LLMs is far more gullible than people. They’re naive and regularly fall for manipulative cognitive tricks that wouldn’t fool a third-grader, such as flattery, appeals to groupthink, and a false sense of urgency. There’s a story about a Taco Bell AI system that crashed when a customer ordered 18,000 cups of water. A human fast-food worker would just laugh at the customer.
Prompt injection is an unsolvable problem that gets worse when we give AIs tools and tell them to act independently. This is the promise of AI agents: LLMs that can use tools to perform multistep tasks after being given general instructions. Their flattening of context and identity, along with their baked-in independence and overconfidence, mean that they will repeatedly and unpredictably take actions—and sometimes they will take the wrong ones.
Science doesn’t know how much of the problem is inherent to the way LLMs work and how much is a result of deficiencies in the way we train them. The overconfidence and obsequiousness of LLMs are training choices. The lack of an interruption reflex is a deficiency in engineering. And prompt injection resistance requires fundamental advances in AI science. We honestly don’t know if it’s possible to build an LLM, where trusted commands and untrusted inputs are processed through the same channel, which is immune to prompt injection attacks.
We humans get our model of the world—and our facility with overlapping contexts—from the way our brains work, years of training, an enormous amount of perceptual input, and millions of years of evolution. Our identities are complex and multifaceted, and which aspects matter at any given moment depend entirely on context. A fast-food worker may normally see someone as a customer, but in a medical emergency, that same person’s identity as a doctor is suddenly more relevant.
We don’t know if LLMs will gain a better ability to move between different contexts as the models get more sophisticated. But the problem of recognizing context definitely can’t be reduced to the one type of reasoning that LLMs currently excel at. Cultural norms and styles are historical, relational, emergent, and constantly renegotiated, and are not so readily subsumed into reasoning as we understand it. Knowledge itself can be both logical and discursive.
The AI researcher Yann LeCunn believes that improvements will come from embedding AIs in a physical presence and giving them “world models.” Perhaps this is a way to give an AI a robust yet fluid notion of a social identity, and the real-world experience that will help it lose its naïveté.
Ultimately we are probably faced with a security trilemma when it comes to AI agents: fast, smart, and secure are the desired attributes, but you can only get two. At the drive-through, you want to prioritize fast and secure. An AI agent should be trained narrowly on food-ordering language and escalate anything else to a manager. Otherwise, every action becomes a coin flip. Even if it comes up heads most of the time, once in a while it’s going to be tails—and along with a burger and fries, the customer will get the contents of the cash drawer.
This essay was written with Barath Raghavan, and originally appeared in IEEE Spectrum.
scalziwhatever_feedI have never been one to care too much about the amount of protein a meal has, but sometimes I see a recipe on Instagram that boasts low calories and high protein and actually looks good, and I find myself tempted to try them out. I mean, if I can eat something healthy-ish and it tastes good, then it’s a win-win, right?
So, after seeing this Buffalo Chicken Hot Pocket recipe, I decided to give it a shot. It seemed like as good a place as any to start with higher protein meals.
Even though the recipe looks long, it’s all pretty simple ingredients, though I did have to go buy quite a few.
So let’s talk about how “quick and easy” it was to make this, how much I had to buy to make it, the time it took, how many dishes it made, and if it actually tasted good.
Diving right in, the first thing was acquiring the ingredients. I shopped at Kroger.
First up, I had to buy a pack of chicken, which ended up being Simple Truth Natural Boneless Skinless Chicken Breast Family Pack for $16.52. I used all this chicken even though it was a big ol’ family pack. Next was Sweet Baby Ray’s Mild Buffalo Wing Sauce for $4.29. I used almost the entire bottle. A block of Philadelphia Reduced Fat Cream Cheese was $3.49. The recipe only needed about a fourth of the block. The recipe calls for a 0% fat Greek yogurt, so I picked Oikos Triple Zero Plain Greek Yogurt, which has zero added sugar, zero artificial sweeteners, and is zero percent fat with eighteen grams of protein (per 6oz serving). I used most of the 32oz container, which was $6.79.
Though I have all-purpose flour, bread flour, and gluten-free flour, I did not have self-rising flour, so I bought King Arthur Unbleached Self-Rising Flour in a five pound bag for $6.29. For the mozzarella, I usually like Sargento’s shredded mozzarella because it’s the only whole milk one I tend to find, but since the recipe specifies a fat free mozzarella, I just went with Kroger Low-Moisture Part Skim shredded mozzarella in the 4-cup size bag for $3.99. I picked Jack’s Special Mild Salsa for my “tomato salsa” which was $4.99 but I have most of the container left over. I also bought Simple Truth Organic Chives for $2.49. And last but not least I bought a Hidden Valley Ranch Seasoning 1oz packet for a whopping $2.39.
I had Daisy brand cottage cheese on hand already, both the whole milk version and the low-fat version, but for this recipe I used the whole milk type since it didn’t specify. Oh, and I used actual whole milk for the quarter cup of fat-free milk it calls for. You’ll just have to live with my substitution.
So, in total, I spent $51.24 on stuff for just this one recipe. I always say you can’t cook dinner without spending fifty bucks, and boy oh boy does that remain true. I swear it’s a literal constant in my life.
Moving on from cost, the first thing to do was to add a bunch of stuff into the Crockpot and let it get cooking. That part was really easy, you just throw the chicken in and add all the spices and whatnot on top, give it a mix and let it cook on high for a couple hours. The only dishes I used for this portion were measuring spoons and a measuring cup. Disclaimer: I did not add the white onion, therefore I saved myself from using a knife and cutting board.
While that was cooking, I blended all the ingredients for the sauce together. I only have a very tiny portable blender meant for protein shakes and smoothies on the go (don’t ask why because I don’t even know), so I had to do it in three or four batches, which meant I mixed everything together in a bowl and then put a couple ladles worth into the blender, blended it and dumped the blended mixture into a separate bowl. Due to my unnecessary steps, you probably will not make as many dirty dishes as I did here. Or as much of a mess on your countertop.
After the sauce was completed, I got to work on the dough. This part was definitely the most time consuming, partially because I decided to be precise and weigh out my ten dough balls to make sure they were perfectly equal. The dough took some work to come together, but after enough kneading, it got there. This portion of the recipe really only took a measuring cup and a bowl, plus the rolling pin to roll out the dough. I set my dough discs aside.
Finally, when the chicken was cooked through, I was very surprised by how much liquid there was in the Crockpot. In the video, when he goes to shred the chicken after its time in the Crockpot, it’s completely dry. I was perplexed why there was liquid in mine, especially when I actually used 100g more chicken breast than the recipe called for. I didn’t want to add my creamy sauce to it while there was so much watery liquid, but I also didn’t want to dump the liquid out of the Crockpot and waste all the flavor that was probably in there.
So, I got to work shredding the chicken to see if it would absorb more as I went. Sure enough, the liquid did reduce quite a bit after the shredding, which took forever and gave my arms a workout. I decided to let the chicken and liquid keep cooking with the lid off for a little bit to see if some of the liquid would cook off or evaporate, and when it finally got decently reduced, I went ahead and added the creamy sauce mixture and all the mozzarella cheese.
It ended up shaping up nicely, and looked like the mixture in the video. All in all, it worked out, it just took extra time. To be fair, the video said cook on high for 2-3 hours and I only did two since the chicken was up to temp.
For the dough discs, I definitely overstuffed the first one, and some of the filling spilled out into the skillet while cooking it. After the hot pocket had been thoroughly browned on both sides, I figured it was done, but when I cut into it, the dough hadn’t cooked all the way through. Though the outside was brown and crispy, the inside was pretty much raw dough. If it had been cooked any longer, though, the outside would’ve burned. I wasn’t sure how to get the inside fully cooked without burning the outside, so this was certainly a predicament.
Plus, my hot pockets were much more oddly shaped than the ones in the video. I couldn’t get a consistent shape and kept second guessing how much filling to put in. It also was pretty time consuming trying to form the hot pockets, and I ended up tearing like two of them. I was definitely frustrated by now, it felt like nothing was working out and I was messing everything up.
After taking a breather and finally eating one of the hot pockets that was cooked through mostly well enough, I am sad to report it was pretty mid. It was fine, but definitely not as good as I had hoped, and definitely not worth fifty dollars and a few hours of work. Though if you consider the fact you get ten hot pockets out of this recipe, it’s only five dollars per hot pocket if you spend fifty on ingredients. I guess that’s not too bad, but I think my feelings of disappointment overshadowed the value of being able to freeze the majority for later.
I will say that there was a pretty decent amount of the chicken filling leftover, whether it’s because I filled the hot pockets the wrong amount or not remains to be seen, but I did like putting the leftover chicken mixture in a tortilla instead. Honestly my main issue with this recipe was the dough. Having the chicken mixture by itself or in a different carb vehicle actually improved my eating experience, I think.
So I would say if you make this recipe, don’t make the dough, and just find something else to put the chicken in, or eat it by itself. Though, there will be less protein in the recipe since the dough was made with protein yogurt. I think that’s worth the trade, though.
Overall, I don’t think I’ll be making this recipe again, but it wasn’t terrible or anything.
Do you like Buffalo chicken? Have you tried Oikos protein yogurt in any of their sweeter/fruitier flavors? Let me know in the comments, and have a great day!
-AMS
icanhascheezburger_feedWhen you're an impending bride, there's an aura of magic that surrounds you and everything you do. Everywhere you go has a luminous glow, and a quiet enchantment to it. Everyday life suddenly feels like you're in the midst of a fairytale. The weeks before a wedding time seems to be held between standing still and going too fast. Seemingly ordinary moments feel imbued with meaning, magic, and messages. Maybe the world knows something important and wants you to know. Your every day is enlightened by the love you have for your soon-to-be spouse. In this liminal space, anticipation is magical in and of itself. You are still fully yourself, still moving through routines and responsibilities but there is an undercurrent of becoming the bride you always wanted to be. As you finalize decisions surrounding your wedding you feel the symbolic weight of everything. These rituals are setting the stage for the rest of your life with your love. Even logistic parts of wedding planning carry emotional weight, because they are tethered to a future that feels suddenly tangible.
Imagine being in that kind of headspace and you stumble upon four lost kittens. It much feel like a fairytale, to see all these kittens curled up together in nature, dozing off happily together as they look up at you with their nuptial eyes. What possible meaning could it have that this woman found four homeless kittens just a mere week before her wedding? In the story, she claims that one of her neighbors must have dropped them off when he went to move, and she severely reprimands his actions. She found the kittens and when they were well enough, she took them in. Even their mom came out and wanted to be adopted. Everyone in the story praises her for taking them, and acknowledges that it's extra special as a wedding gift.
icanhascheezburger_feedWe just want to make sure, since it's a known fact that every cat pawrent's life is controlled by their fluffy feline friend they decided to let into their house, overthrow whatever house government was in place, and start their royal meownarchy, ruling the house with an iron paw. You know, your feline overlord who dictates your every second in the daily schedule of the house and beyond.
So, did they give you purrmission? We really don't like to overstep on cats' schedules, and if you're a cat pawrent looking for some funny cat content online, we have to make sure - did you cat allow you this time? Or your cat ordered you to look for some silly cat memes for their own amusement? We just want to know that whatever you're doing right now is only in servise of the feline emperor who's actually in charge.
We don't want to imagine the punishment of not following your feline's orders to the last meow. The consequences are dire. Your cat won't show you their belly for a week, and they will push every last mug off the counter. So, what did your cat say? If they allow you this free time, feel free to look at this cute cat collection of comedy, just as you intended in the first place.
icanhascheezburger_feedAll of us will have to deal with the rainbow bridge sooner or later. It's the circle of life, and as sad as endings are, they are also the doorway to new beginnings. We're not saying that to tell you that it should be easy, far from it. Every family member, including your other pets, feels the loss of a cat, and it hits some of them harder than others. The cat in our story today was devastated after losing his senior sibling. Months passed, and his condition hadn't improved. But then, they were alerted to a local woman who had a litter of kittens she was giving away, and they decided it was finally time to start over again.
The new kitten healed the cat's broken heart, as his pawrents walked in on them cuddling in the sweetest way. They were purrfectly embraced. You could purractically feel the joy in his tiny, furry body over having another cat to love and share everything with.
It's not uncommon to see cats sleeping piled on top of each other, but these two are doing it differently. They are literally spooning, with the now older brother pawsitively enveloping the baby kitten and protecting them with his paws. It's quite pawssibly the cutest thing we've seen in a long time.
Everyone heals from grief differently. Some need to focus on something new, and others need time to process. But what we can all agree on is that getting a new cat will never erase the old one. His memory will live on furrever in all of our hearts, which are now filled with joy again with the excitement of a new beginning.
snopes_feedpattrosesonianaomikritzersovayspatch's fifty-first I took him to Porter Square Books and on the roundabout way home we collected dinner from Il Casale. It started to snow on the way back, the light salting flakes of an all-day deep-freeze. I have my fingers crossed for an Arctic explosion this weekend.
![[community profile]](https://www.dreamwidth.org/img/silk/identity/community.png){kind=small}
threesentenceficathon. WERS played Dave Herlihy's "Good Trouble" (2025) and I had to get home to trace his voice to Boston's own post-punk O Positive. I wish I could call the hundred-year tides against the people who have no right to the streets of my grandparents' city. Failing that, it still matters to be alive.aurumcalendulacorvidology:icanhascheezburger_feedThe world as we know it changed when cats were no longer our foes, but our independent domestics who sometimes hang out with us because we give them little crunch treats. They know how to survive, and if that means they have to stomach hooman cuddles every so often, then so be it. You would do it too, for a Churu.
Hissterical feline funnies have a habit of making us chuckle at the worst pawssible times. Where were you the last time you came across a hilarious cat meme in the wild, and you had to attempt to stifle a laugh, only to fail miserably? Since I scrub through and consume cat content for a living, the open-concept office hears my laughs day in and day out, unfortunately for them. Who cares… A cat meme will do what a cat meme is going to do. We hoomans are all lowly subjects to the royal feline gods and goddesses, and what we do with the hissterical content they give us is no one else's business. Especially those who do not prefer fluffy felines…
Though we're their subjects, cats do their very best (unintentionally) to put a smile on our faces when the world seems blue. Don't believe me? Scroll below, you one be disapawnted.
landofnowhereicanhascheezburger_feedYou can say their name five times in a row and get nothing in return. No glance. No ear twitch. Not a single acknowledgment that sound even happened. But the moment a bag rustles or a bowl clinks in the kitchen, suddenly they're wide awake and extremely present. Cats hear everything. They just prioritize.
They'll sit right next to you and still ignore you. Close enough to feel involved, far enough to avoid responsibility. You reach out to pet them and they act like you suggested something unreasonable. A slow turn away. A sudden interest in grooming. Conversation over.
Ignoring hoomans is also strategic. Ask for attention and they vanish. Decide you're busy and no longer care? Now they're back, sitting on your keyboard or blocking your view like they've been there all along. The timing is never random.
Food, however, changes everything. Offer snacks and suddenly you exist again. Eye contact is made. Bodies move with purpose. Affection appears out of nowhere. It's amazing how fast indifference turns into enthusiasm when treats are involved.
Still, every so often, they surprise you. A quiet sit nearby. A gentle head bump. A slow blink that feels intentional. Those moments hit harder because they're rare. Cats may ignore their hoomans most of the time, but when they choose to engage, it feels special. And honestly, that's part of the charm.
ursulaglowingfish asked:The Golden Age of published science-fiction was more or less from 1955 to 1975 (lets say). Why did it end when it did? Do you think that science-fiction (or fantasy) published after 1975 was different, or do you just think it had less ability to become part of the "canon"?
glowingfish's, I'm guessing), and it makes sense that the stories they considered formative would seem quasi-canonized to our generation.pattrosemathhobbit - Word of the Daybruce_schneier_feed - Why AI Keeps Falling for Prompt Injection Attacksscalziwhatever_feed - A Very Mid Attempt To Incorporate More Protein Into My Dinnericanhascheezburger_feed - Soon-to-be bride discovers four homeless kittens, finds out her neighbor left them when he moved soicanhascheezburger_feed - 28 Funny Feline Memes for People Whose Lives are Completely Cat-Controlledicanhascheezburger_feed - This cat was crushed after his senior sibling crossed the rainbow bridge, until his pawrents adoptedsnopes_feed - Did Italian TV show claim 'Goodfellas' image depicted real Iranian family before 1979 Islamic Revolusnopes_feed - Did 'historic wind event' leave massive snow bank on Wyoming's Devils Tower?pattrose - Not quite the 365 days question meme January 22icanhascheezburger_feed - It's all just fashion anyways.sonia - Links: Minnesota general strike, and how to helpsovay - No, I'll build a cute flower borderaurumcalendula - January Talking Meme - Dream place to live?icanhascheezburger_feed - Free Willicanhascheezburger_feed - 22 Cute Cat Memes for Those in Need of Some Furry Purry Feline Comedy This Winter (January 21, 2026)landofnowhere - wednesday books, in which women guide the plotcelli - how many tags can I cram into this entry?icanhascheezburger_feed - 21 Relatable Memes of Cats Majoring in Selective Hearing and Ignoring Their Pawrentsursula - January journal: science fiction periodizationpattrose - 2026 SNowflake Challenge #11dancing_serpent