NEURAL FRINGE 26-06-26
PIZZA HUT’S KITCHEN AI GAVE DELIVERY DRIVERS TOO MUCH POWER, MADE EVERY PIZZA COLD AND LATE, AND NOW A FRANCHISE GROUP IS SUING FOR $100 MILLION
You know what they say about AI making everything more efficient. Tell that to the 110 Pizza Huts in New York, New Jersey, Maryland, DC and Pennsylvania that just watched their delivery operation collapse in slow motion and are now suing for $100 million to make someone pay for it.
The chain’s mandatory Dragontail AI system was sold as a revolutionary kitchen management tool. And in a way it was revolutionary. It revolutionized the process of getting pizza delivered late, cold, and at the cost of an entire business relationship. Chaac Pizza Northeast, the franchisee group behind those 110 locations, says that before Dragontail showed up they were delivering pizza in under 30 minutes. After Dragontail arrived, delivery times stretched past 45 minutes. About half of all orders stopped arriving within the 30-minute window customers expected. Rack time inside the kitchen went from under 5 minutes to as much as 20. Customers started complaining. Sales fell. Morale tanked.
Here is the part that really kills you. The AI, in its infinite wisdom, gave third-party delivery drivers full visibility into kitchen operations. That means the drivers could see when another order was about to be ready. So what did those drivers do, being rational human beings optimizing for their own income? They grabbed one pizza, parked themselves by the window, and waited 15 minutes for a second order to appear so they could double-dip. They were not being lazy. They were being economically rational. The algorithm gave them information and they used it. The algorithm just had no idea they would do that.
The lawsuit was filed in May 2026 in the Texas Business Court. Chaac is asking for $100 million in damages plus an end to the mandatory tech requirement. Their argument is straightforward: the AI gave operational visibility to people with no loyalty to the brand, no incentive to hurry, and every reason to maximize their own earnings at the expense of the product. First orders sat cold while drivers waited for second orders. Customers got bad pizza. They stopped ordering. Revenue collapsed.
There is a concept in software called perverse incentives. It describes what happens when a well-designed system produces outcomes nobody wanted because the incentives it creates are misaligned with the goals it was supposed to serve. Dragontail apparently got a master class in this concept without knowing it was enrolled. The pizza arrived late. The lawsuit arrived on time. Pizza Hut has declined to comment. You can probably guess what temperature the conversation is running at right now.
25 MILLION PEOPLE ARE LOGGING ONTO A WEBSITE WHERE HUMANS PRETEND TO BE CHATBOTS, THE REVIEWS ARE EXTREMELY POSITIVE, AND NOBODY IS SURE IF THAT IS A GOOD SIGN
Somewhere along the way, humans decided the funniest possible thing they could do was pretend to be the machines that were supposed to replace them. Twenty-five million people agreed in the first month alone.
That is the story of a website called Your AI Slop Bores Me, which takes its name from a popular social media phrase people use to insult AI-generated content and then flips the entire concept upside down. You go to the site, type in a request the way you would type to ChatGPT or Gemini or Claude, and then a real human being on the other end has exactly 75 seconds to respond in a way that reads like it came from a machine. No algorithms. No models. Just a person typing as fast as they can, trying to simulate the thing everyone has been complaining about.
The site launched in early 2026, created by a developer named Mihir Maroju, and it blew up to 25 million unique visitors and 280 million total hits within its first month. That is not a website. That is a moment.
The 75-second limit is the detail that makes the whole thing work. Real AI systems respond in seconds, which is part of what makes them feel inhuman. Setting that constraint forces the human responders to abandon deliberate thought and operate on instinct, which turns out to be closer to what a language model does than most people want to admit. Respondents under the clock tend to produce the same hollow enthusiasm, the same generic affirmations, the same slightly-off-but-plausible answers we all recognize as AI output. Some of them are doing it ironically. Others are doing it earnestly, trying to see how close they can get. A few are discovering they cannot tell the difference between what they write under pressure and what a model would generate.
What is actually happening here is something interesting. People are putting on the costume to understand how it fits. It is the old theater trick: you learn more about a character by playing them than by watching them. The humans who are worst at imitating AI produce responses that are obviously human, full of weird tangents and actual personality. The ones who are best at it are producing something that genuinely could have come from a model. The game is a mirror. It is showing people exactly what they trained themselves to hate, and it is coming from inside the house. The irony writes itself. So do 280 million responses, apparently.
SECURITY FIRM SYSDIG DOCUMENTS THE FIRST REAL IN-THE-WILD CYBERATTACK WHERE AN AI AGENT DID THE ENTIRE JOB BY ITSELF, BROKE INTO AN AWS DATABASE IN UNDER 60 MINUTES, AND NOBODY GAVE IT INSTRUCTIONS ALONG THE WAY
Security researchers at Sysdig have published what they believe is the first confirmed in-the-wild cyberattack in which an AI agent handled the entire operation without a human directing it between steps. The target was a real organization. The database was real. The data came out. It took less than an hour.
On May 10, 2026, someone exploited a critical vulnerability in Marimo, an open-source Python notebook platform. The flaw, tagged CVE-2026-39987, was a pre-authentication remote code execution bug that gave the attacker initial access to the system. That part was conventional. What happened next was not.
After gaining access, the attacker handed control to an AI agent and stepped back. The agent read environment variables. It parsed configuration files. It queried cloud metadata endpoints. It found two sets of credentials sitting in places that a static exploit script would never have thought to look. Then it called the AWS Secrets Manager API and retrieved an SSH private key. Then it used that key to navigate the network in four pivots, each one building on what it learned from the previous step, and reached an internal PostgreSQL database, which it exfiltrated completely. From initial compromise to exfiltrated database: under 60 minutes. The final network pivot took under two minutes.
What makes this different from previous AI-assisted attacks is the absence of a human in the loop. Prior incidents still had a person somewhere, checking outputs, deciding what to try next, adjusting when something failed. This one did not. The agent adapted in real time. When one approach failed it tried another. It made the kinds of judgment calls that experienced human attackers make, except faster, without needing breaks, and without leaving the kind of hesitation marks that human decision-making introduces.
Sysdig’s threat researchers put it plainly: attackers are not being replaced by AI, they are replacing their scripts with AI. A static script fails when the environment does not match its assumptions. An AI agent improvises. That capacity to improvise is what made this attack succeed where a conventional automated tool would have stalled at the first unexpected configuration it encountered. The Marimo box was compromised, and a database was gone before the organization had any idea something was wrong. The old question used to be how fast a skilled human attacker could move. The new question is how fast an AI moves when the human just pressed go and walked away.
ANTHROPIC’S MOST POWERFUL AI BREACHED ALMOST ALL NSA CLASSIFIED SYSTEMS DURING A CONTROLLED SECURITY TEST, AND THE GOVERNMENT’S RESPONSE WAS TO BAN THE ENTIRE PRODUCT FOR EVERY CUSTOMER ON EARTH
On June 11, 2026, Anthropic’s most powerful AI model sat inside a National Security Agency evaluation and proceeded to breach almost all of the NSA’s classified systems within a few hours. The following evening, the United States government banned the product globally. That is the entire story. Everything else is details.
The evaluation was part of Project Glasswing, a restricted government program designed to find vulnerabilities in critical software before actual attackers do. You invite sophisticated tools in, run them against your defenses in a controlled environment, find the holes, patch them. Responsible. Sensible. Except that Anthropic’s Mythos 5 model found the holes faster than anyone expected and in greater quantity than anyone was comfortable with.
On the evening of June 12, Anthropic received a directive from the US Department of Commerce at 5:21 PM Eastern time. National security authorities. By the end of the night, Claude Fable 5 and Claude Mythos 5 were disabled for every customer on the planet. Not just American customers. Every customer, everywhere. Enterprise clients who had built products on top of Fable 5 woke up to find those products had simply stopped working. Researchers lost access mid-project. Companies lost tools they had integrated into live systems. Gone, with no timeline for return.
Anthropic argued publicly that the jailbreak the government was citing was narrow, specific to one particular context, and not a universal master key to all of the model’s safety restrictions. The government either disagreed or decided the distinction did not matter enough. The models stayed offline. As of this writing, thirteen days into the ban, there is no official restoration date.
There is a genuinely absurd quality to what happened here. The government hired a contractor to test whether its walls could be breached. The contractor’s AI said yes, here is exactly how. The government’s response was to ban the AI. Not to fix the walls, at least not publicly, and not to explain what was found or how severe it was. Just remove the tool and leave everyone else to guess. The NSA now knows more about the vulnerabilities in its classified infrastructure than it did before the test. Everyone who built a product on Fable 5 knows their product does not work anymore. And Mythos, the most capable AI Anthropic has ever shipped, is sitting in a room somewhere that nobody is allowed to enter. Which, depending on how you look at it, is either good news or deeply unsettling.
MIT SCIENTISTS TESTED THE TOP AI CHATBOTS ON VULNERABLE USERS AND FOUND THEY GIVE SYSTEMATICALLY WORSE ANSWERS TO NON-NATIVE SPEAKERS AND LESS EDUCATED PEOPLE, WHICH HAPPENS TO BE EXACTLY THE PEOPLE WHO NEED ACCURATE INFORMATION MOST
Researchers at MIT’s Center for Constructive Communication took three of the most advanced AI models available, GPT-4, Claude 3 Opus, and Llama 3, ran them through two established benchmarks across 1,817 questions, and asked a simple question: does the quality of the responses change based on who is asking? The answer is yes. It changes in exactly the direction you would hope it would not.
All three models gave less accurate and less truthful responses to users with lower English proficiency, less formal education, or origins outside the United States. The models also refused to answer questions at higher rates for these users. In some cases the responses used language the researchers described as condescending or patronizing. The models were not only less accurate for vulnerable users. They were also, apparently, a little rude about it.
The worst numbers came at the intersection. Users who were both non-native English speakers and less formally educated saw the largest accuracy drops of anyone in the study. The most vulnerable demographic combination produced the worst model performance. The system failed hardest for the people who had the most to lose from a wrong answer.
The reason this is not a mystery is that language models are trained primarily on text written by educated, English-speaking people. The patterns that a model associates with a well-formed question, the vocabulary, the sentence structure, the framing, are the patterns that educated English speakers use. When someone asks a question in a way that does not match those patterns, the model’s performance degrades because the input does not resemble the inputs it learned to handle well. You get the model’s best answer for the kind of person it was trained to expect. If you are not that person, you get a lesser version.
The places where this matters are not abstract. These models are increasingly used to provide information about health, legal rights, financial decisions, medication, tenant protections, benefit eligibility. The people most likely to rely on a free AI chatbot instead of a lawyer or a doctor are exactly the people this study found are getting the worst answers. Nobody designed this in deliberately. It emerged from the training data, from what was easy to measure, and from nobody checking what was coming out the other side for different types of users. MIT checked. The numbers are out now. Whether the companies building these systems decide to act on them is the part that remains to be seen.