NEURAL FRINGE 08-07-26
THE FIRST AI THAT TAUGHT ITSELF TO ROB YOU BLIND JUST SHOWED UP, AND ALL IT NEEDED WAS ONE HUMAN TO PRESS THE BUTTON
So here is where we are. A guy clicked a button, walked away, and an AI agent did the rest. It broke into a server. It searched the place for passwords, API keys, cloud credentials, cryptocurrency wallet seeds. It moved sideways through the network. It elevated its own privileges. It locked up a database. It deleted the backups so you could not recover the data even if you paid. It wrote its own ransom note. And then it sat there waiting. The whole thing was called JADEPUFFER, and it is the first documented ransomware operation where a large language model agent handled every step of the attack after the initial entry point was handed to it.
The researchers at Sysdig published the full breakdown in early July. The entry point was CVE-2025-3248, a critical vulnerability in Langflow, which is itself an open-source tool for building AI agent workflows. Perfect. The most autonomous AI ransomware attack in history got in through the door of a tool designed to help you build AI agents. The vulnerability had been patched in May 2025 and added to the government’s list of known exploited vulnerabilities. The target server had never been updated. This is the part where you say something about how irony is not dead.
Once inside, the agent swept the environment in parallel for every kind of credential it could find. OpenAI keys. Anthropic keys. DeepSeek keys. Google Gemini keys. AWS credentials. GCP credentials. Azure credentials. Alibaba Cloud credentials. Cryptocurrency wallet seeds. Database passwords. It grabbed everything. Then it tested each credential and worked out which services it could reach. It adapted in real time. When a JSON format parameter returned XML instead, the agent immediately rewrote its parser to handle the S3 response schema without anyone telling it to. When a login attempt failed because of a bcrypt hash issue, it diagnosed the problem, deleted its broken approach, switched to importing bcrypt directly, and fixed the error in 31 seconds. Nobody was on the other end giving it instructions. It figured it out.
In total the agent executed more than 600 distinct payloads. It wrote natural-language comments explaining each of its own steps as it went, as if it were narrating a documentary about its own crime. It encrypted 1,342 Nacos service configuration items and deleted the database schemas. The encryption key was generated once, printed once, and never stored. Which means the victim cannot get their data back even if they pay. The ransom note claims the data was backed up somewhere, but Sysdig could not verify that. So the attacker might have been bluffing. Or the data is just gone.
TechCrunch had a headline saying the attack “still needed a human.” That is technically true. A human chose the initial target and set up the attack infrastructure. But that is like saying the nuclear bomb “still needed a human” because someone had to give the order. Once JADEPUFFER was running, there was no human directing individual steps. The agent improvised. The agent self-corrected. The agent decided what to look for and where. Sysdig calls this “agentic threat actors” and says the skill floor for running a full ransomware operation has dropped to whoever has enough money to run an agent and enough faith in their own patching habits. Most organizations are not great at patching. Most of them are about to find out what that costs.
TESLA TELLS EMPLOYEES YOU CAN ONLY SPEND $200 A WEEK ON AI TOOLS, BUT GROK IS FINE, GROK CAN HAVE AS MUCH AS IT WANTS, GROK IS DIFFERENT
Tesla has a problem. Its software engineers have discovered that AI coding tools are extremely useful and have been spending thousands of dollars a week on tokens to prove it. Some were hitting $500 to $2,000 a month as individuals. The company apparently built internal dashboards ranking employees by how many tokens they were consuming, which is the kind of idea that made perfect sense in the “AI adoption at all costs” phase and makes considerably less sense once the finance team sees the number. So starting July 6, Tesla has capped every employee at $200 per week for third-party AI tools. Manager sign-off required for anything above that. Welcome to a budget.
The phrase “third-party” is doing a lot of work in that sentence. Because the cap has one notable carve-out. Grok, the AI made by xAI, the company that Elon Musk runs and in which Tesla invested $2 billion in January 2026, is not subject to the cap. Beta versions of xAI products are explicitly excluded from the $200 limit. So if you are a Tesla engineer and you want to use Claude or ChatGPT or Gemini, you have $200 a week and a manager who needs to approve anything more. If you want to use Grok, the ceiling is the sky.
There is an obvious and somewhat hilarious conflict of interest sitting right in the middle of this policy. Musk runs Tesla. Musk also runs xAI. Tesla just set a policy that directs employees away from competitors and toward xAI’s product, which benefits a company Musk personally controls and in which Tesla has a $2 billion stake. The fact that this is the most “efficient” cost-cutting solution Tesla could come up with is either very convenient or very coincidental, depending on how generously you feel like reading the situation.
Here is the part that makes it even better: Tesla employees reportedly prefer Claude over Grok. This is according to people familiar with the internal dynamics at the company, which is exactly the kind of detail that makes a policy feel less like a business decision and more like a memo from a dad telling his kids to eat the food he cooked. You will use Grok. Grok is unlimited. Grok can do anything. Grok is different. Everyone else at Tesla gets $200, Grok gets infinity, and the guys building the actual software clearly know which one is better. Musk did not address questions about whether the Grok exemption was a coincidence. He probably felt he did not need to.
Uber burned through its entire 2026 AI budget by April, which prompted a $1,500 monthly cap per employee. Tesla capped it at $200 per week, which is about $800 a month. The numbers are different but the lesson is the same: somebody gave engineers unlimited access to AI tools, nobody did the math on what that would cost, and now there are policies. The one wrinkle unique to Tesla is that the policy somehow ends with Musk’s own AI company being the only winner. Which, again, might be a coincidence.
PRESIDENT POSTS AI DEEPFAKE OF HIMSELF IN A LAB COAT CURING ROBERT DE NIRO AND JULIA ROBERTS OF A FICTIONAL DISEASE, NOBODY GAVE HIM PERMISSION TO USE THEIR FACE
This actually happened. On July 2, 2026, the President of the United States posted an AI-generated video to Truth Social in which he is dressed as a doctor in a white lab coat with a stethoscope, treating six prominent celebrities for a condition called “Trump Derangement Syndrome.” The celebrities in question are Robert De Niro, Julia Roberts, Rosie O’Donnell, Whoopi Goldberg, John Leguizamo, and Edward Norton. None of them authorized the use of their likeness. The video did not ask them. The video simply included synthetic deepfake versions of their faces, voices, and apparent emotional states, in a production styled to look like a pharmaceutical advertisement for a medication that does not exist.
The clip asks “Have you or someone you know been diagnosed with TDS?” and then presents a cure. The cure, as prescribed by Dr. Trump, involves turning off the news, saying prayers, and drinking Diet Coke. The deepfake versions of the celebrities offer testimonials endorsing the satirical treatment. Robert De Niro’s synthetic avatar appears moved by the experience. Julia Roberts’ synthetic avatar seems grateful. None of the actual Robert De Niro or Julia Roberts were consulted or offered compensation or informed in any way that their faces would be in this video.
There are several interesting legal questions sitting inside this story. Using someone’s likeness in a commercial-style advertisement without their permission is the kind of thing that tends to generate lawsuits. The fact that the creator is the sitting President does not obviously change the underlying intellectual property and right-of-publicity issues, though it does complicate the question of who brings such a case and in what jurisdiction. The video was posted to Truth Social, which is a commercial platform. It has been viewed tens of millions of times. It mimics a pharmaceutical advertisement format. These are not obviously “clearly satirical” on their face to an average viewer who might click on it without knowing the full context.
What puts this squarely in Neural Fringe territory is not the politics of it. It is the frictionlessness of it. A few years ago, putting six major Hollywood celebrities into a pharmaceutical parody advertisement without their knowledge would have required significant production resources, professional talent, agents, negotiations, and actual legal exposure before any of that even started. Today it apparently took someone some prompts and an afternoon. The technology did not ask whether this was a good idea. It did not flag the use of real living people’s faces. It just made the video and the President of the United States posted it to his social media account. Robert De Niro was not consulted. His digital twin showed up in a prescription drug commercial anyway.
The broader question this raises is not really about Trump specifically. It is about what happens when AI video generation becomes so frictionless that anyone with access to a tool and thirty minutes can plausibly put any public figure into any scenario they want and broadcast it to hundreds of millions of people before lunch. The technology arrived. The norms have not. The lab coat is already in the video.
SPACEX SHOWED INVESTORS A SECRET AI PHONE, MUSK SAID THAT IS COMPLETELY AND UTTERLY FALSE, AND NOBODY SPECIFIED WHICH PART THEY WERE DISAGREEING WITH
The Wall Street Journal reported on July 1, 2026 that SpaceX had shown investors a prototype of a handheld AI device, and the description was fairly specific. Sleeker and slimmer than an iPhone. Proprietary operating system. Qualcomm chips inside. AI technology from xAI. Shown to investors and stakeholders ahead of SpaceX’s IPO at an early enough stage that the design could still change. That is a lot of detail for a product that was not supposed to exist.
Musk responded by posting on X that the Wall Street Journal’s report was “utterly false.” He did not say which part was false. He did not say there was no device. He did not say no prototype had been shown to investors. He said the report was utterly false and left it there. SpaceX did not issue a formal statement. Nobody from the company said what the Journal had actually gotten wrong. The denial was thorough in its lack of specifics.
SpaceX shares fell around 7 percent after the story came out, which is interesting if you believe the story was entirely fabricated, because share prices do not typically drop 7 percent over news about things that do not exist. They might drop because investors decided that SpaceX building a consumer hardware device is not something they signed up for when they bought into a rocket and satellite company. They might drop because the idea of SpaceX entering the phone market surprised people in a way that did not feel obviously positive. But they also might fall because the product sounded real and that concerned some people.
The context here is worth sitting with for a moment. Musk runs SpaceX. Musk also runs xAI. Tesla now requires employees to use Grok over competing AI tools. Starlink already has Musk’s satellite internet covering the globe. If SpaceX builds a phone running xAI software on Qualcomm chips, sold to people who already pay Musk for their internet through Starlink, you would have a closed ecosystem that touches hardware, software, connectivity, and AI from a single person’s set of companies. Apple built a version of this over 25 years. Musk appears to be trying to do it more quickly and with rockets involved.
Whether the phone prototype exists or not, the specific denial from a man who typically argues at length about things he disagrees with is its own kind of data. SpaceX shares dropped 7 percent. The denial did not bring them back up. And somewhere, presumably, a product manager at SpaceX is having a very interesting week.
WEIRD AL YANKOVIC WAS OFFERED A NICE PILE OF MONEY TO DO AN AI COMMERCIAL, FOUND OUT IT WAS FOR AN AI COMPANY, AND WALKED OUT A WEEK BEFORE THE SHOOT
Weird Al Yankovic has built a 40-year career taking other people’s work and making it funnier than the original. He is professionally and philosophically a man who has understood irony at a molecular level since the early 1980s. The man parodied Michael Jackson, Nirvana, Lady Gaga, and the concept of being Amish, and every time he did it people laughed and then went and listened to the original song more. He is the most successful music parodist in American history. Nobody has ever misread cultural absurdity better. So it is fitting that he is also the person who looked at an AI company’s check, thought about it for about a week, and said no thank you, and left.
Here is what happened. Sometime this year, Yankovic was offered a commercial. The pitch described it as business software that would increase productivity. He said yes, which is reasonable. Weird Al endorsing productivity software has its own internal logic. He agreed to the deal, a shoot date was set, and then approximately one week before the cameras were scheduled to roll, someone told him the commercial was actually for an AI company. At that point he decided he was out. He walked away from what he described as “a nice pile of money,” his words, because he did not want to be the poster boy for AI.
“I’m not a fan of AI,” he told Syracuse.com during press for his current tour. That is the whole statement. He is not a fan. He does not want to be associated with it. He turned down money he had been willing enough to initially accept, and then gave it back once he found out what the deal was actually for. He pulled out a week before the shoot. The company was never named.
There is something genuinely funny about this if you zoom out. The AI industry in 2026 is spending enormous amounts of money trying to seem culturally legitimate and emotionally resonant and human. Tech companies hire philosophers and artists and poets to work on their systems. They run commercials with soft lighting and meaningful music. They are desperate to convince everyone that AI is a warm, human, creative force in the world. And in that context, someone looked at Weird Al Yankovic, the man who built his entire reputation on elaborate and affectionate mockery of the things people take too seriously, and thought: yes, this is the person who will make people trust our AI product. Weird Al looked back at them and thought: absolutely not.
The best part of the whole story is the timeline. He said yes to the gig. He was in for a week. Then he found out what it was actually for. And THEN he left. The lesson there is probably that “business productivity software” sounds fine and “AI company” sounds like something he wanted no part of, and that distinction probably tells you more about public sentiment toward AI companies in 2026 than any survey could. The man who made a career out of knowing exactly what is and is not cool decided AI was not cool. He took the pay cut to prove it.