Mad Cow. The AI, Privacy, and Security Weekly Update.
EP 316.
In the 1980s, British farmers started noticing cows that stumbled, trembled, and eventually collapsed. The cause, researchers later showed, was a prion disease that spread through cattle feed.
To save money, producers had been grinding up the remains of other cattle and feeding the result back to herbivores.
The animals were eating their own kind, secondhand, and the disease circulated through the herd until the feed itself was the problem.
Keep that picture in mind this week, because the AI industry is starting to look a lot like that feed lot.
Machines write stories, and those stories end up in the training data for the next generation of machines. Models learn from the output of other models. Hackers use AI to decide which victims are worth robbing. Criminals steal AI access and run up bills for companies that never signed up for the job. Satellites are being built to run AI in orbit. Phone carriers keep years of records, cars sell driving habits, and exchanges lose billions through laptops that trusted too much. Meanwhile, the UK is switching off its landlines in 100 days.
The question underneath all of it is simple.
When a system keeps consuming its own leftovers, what does it end up producing, and who gets sick?
Welcome to Mad Cow.
Global: Elias Could Be Proof of AI's Mad Cow Disease
Ask AI to write a story, and you might meet Elias, a lighthouse keeper who apparently never gets a day off.
Cornell researchers examined 20,000 stories generated by ChatGPT, Claude, Gemini and OLMo. More than half featured a lighthouse. Nearly 90% included at least one of 11 recurring words, including "keeper" and the names Elias, Mara and Elara. In one test of Anthropic's Claude, 56% of the stories shared the exact same title: "The Lighthouse Keeper's Secret." Elara was even more overused before Elias took over: the linguistics site Namerology named it the 2025 AI Name of the Year. Different chatbots, suspiciously familiar imaginations.
The researchers weren't requesting nautical fiction. They used variations of "tell me a story," yet familiar characters and settings kept surfacing. One possible source is WildChat, a collection of human conversations with ChatGPT used to train other models. Researchers found Elias and lighthouse stories there, too. That's a clue, not a conviction: they haven't established exactly why these particular ingredients keep coming back. One explanation reportedly points to human feedback. These tropes appear to be overrepresented not in classic literature or web text, but in the preference datasets used to train AI, where human raters repeatedly upvoted safe, cozy storybook themes over edgier material.
Think of it as a digital parallel to mad cow disease, with an important distinction: this is a feedback loop, not an infection. Mad cow disease spread through feed containing infected animal material. With AI, yesterday's machine-generated writing can become tomorrow's training material, carrying familiar patterns into another generation. The mechanisms differ, but the comparison lands: trouble can grow when a system keeps eating its own leftovers.
Separate research shows that indiscriminate training on AI-generated material can cause "model collapse," where uncommon details disappear and outputs lose variety. Cornell's recurring lighthouse keeper doesn't prove that's happening here, but it illustrates why researchers worry about a shrinking creative menu. Polished writing isn't necessarily original writing, especially when several machines keep serving the same special.
So what's the upshot for you?
Use AI to help shape your writing, but bring your own reporting, experiences, and specific details to the draft. If everyone else is reheating the same machine-made leftovers, your advantage is bringing something nobody else has tasted.
If AI is recycling its own stories, the people on the other end are being recycled too. Hackers have found a way to use machine-written analysis to decide which of us is worth robbing in the first place.
Global: AI Is Helping Hackers Decide How Much You're Worth
Hackers have a new use for AI: figuring out which victims are worth robbing. Researchers at Cisco Talos uncovered a Windows malware strain called Volta-Autodoxer that doesn't just steal information.
It uses large language models to analyze everything it collects and build detailed profiles of its victims. The haul includes passwords, credit cards, cryptocurrency wallets, browser sessions, Wi-Fi credentials, and even screenshots.
The malware doesn't stop at what's stored on your computer. It uses stolen access to Gmail, Amazon, DoorDash, and Uber Eats accounts to dig through receipts and purchase histories. Delivery apps are a particular focus. Uber Eats and DoorDash receipts can reveal a victim's physical address, sleep patterns, and late-night eating habits. From there, it can confirm your home address, identify your employer, understand your habits, and estimate what you own. Essentially, it's building a personal intelligence dossier on you. Eeek!
The stolen information is then sent to an AI model instructed to act like an intelligence analyst. It produces a report on what you're worth, ranks your potentially valuable assets, and assesses how good your security practices are.
The malware initially uses Anthropic's Claude, but if the model refuses the request, it can switch to alternatives from Moonshot AI and Google. One AI saying no doesn't necessarily stop the operation.
There is an important qualification: Cisco Talos has analyzed the malware but has not established whether it has been used in actual attacks. Still, the design shows how AI can make traditional cybercrime more efficient. Instead of criminals manually sorting through thousands of stolen records, AI can do the investigative work and point them to the most profitable targets.
So what's the upshot for you?
Time to stop thinking of your email account as just an inbox. It's increasingly the master key to your digital life. Use unique passwords, enable phishing-resistant MFA wherever possible, secure your email account first, and regularly review connected applications and active sessions.
If hackers can get into your email, they may not just steal your identity. They can use AI to figure out what else you're worth stealing.
Once criminals know what you're worth, they look for ways to make you pay for their operation. Sometimes the bill for their AI habit lands on someone else's account entirely.
Cybercriminals have found a way to make the AI boom even more expensive: steal a company's access and stick it with the bill. Google's Mandiant investigators say fraudulent AI usage can exceed $100,000 a day. Some victimized organizations accumulated over $1,000,000 in unauthorized charges before noticing the compromised keys and shutting them down, The Register reports.
The scam is called LLMjacking. Thieves grab digital keys that let software access paid AI services, then use that access themselves or sell it to others. In one incident, an attacker made more than 500,000 AI requests in under eight hours using stolen keys to power a service called Jet Night.
Google suspects Jet Night's bargain pricing depends on somebody else paying the actual costs. It advertised unlimited access to advanced AI models for roughly $60 a month. Investigators also noted repeated outages, likely because victims discovered the theft and canceled the stolen keys. Cheap AI apparently gets less reliable when the unwilling sponsor notices.
The tricks go beyond stealing keys. One fake membership portal persuaded users to paste login-session data that let criminals bypass passwords and multifactor authentication. Google recommends limiting account permissions, regularly replacing access keys, and checking software code for exposed credentials.
So what's the upshot for you?
Consider AI access keys like credit cards, never paste login-session data into a stranger's site, and be skeptical of unlimited premium AI at bargain prices. The smartest thing your AI does this month might be refusing to finance somebody else's startup.
Stolen AI access is someone else's bill, and the industry's answer to its own hunger is to build bigger kitchens. Google's latest one isn't on the ground at all. It's in orbit.
LEO: Google's New Data Center Is Out of This World
Google has taken another step toward putting AI data centers in space. Its first experimental satellite, part of a project called Suncatcher, is operating in orbit after launching earlier this month aboard a SpaceX rocket.
The satellite carries four of Google's specialized AI processors, called TPUs, and has already communicated with engineers on Earth. Its first assignment was to write a haiku about being in space. The Gemma model running onboard sent back its first downlinked message, received over Svalbard, Norway:
Blue marble below
Data streams in silent grace
Starlight feeds my core.
There is a more serious purpose behind the experiment. Google wants to build networks of satellites that can process AI requests directly in orbit, rather than relying entirely on enormous data centers on Earth. The attraction is constant solar energy, fewer restrictions on land, and potentially lower operating costs as the technology develops.
The engineering problems are significant. The satellite generates about one kilowatt of electricity, enough to run only one of its four processors at a time, for roughly 20 minutes per orbit. Engineers are also testing how space radiation and extreme temperatures affect the chips. Radiation can flip individual memory bits, potentially corrupting data without triggering conventional monitoring systems.
If the experiment succeeds, Google plans to launch two additional satellites in 2027. These will test higher power capacity and laser-based communications between satellites, allowing them to work together on larger computing tasks. The long-term ambition is a network of interconnected orbital data centers, potentially becoming a new foundation for AI infrastructure during the 2030s.
So what's the upshot for you?
The race to build AI infrastructure is moving beyond traditional data centers, and that could eventually change the economics of computing, energy and cloud services.
Keep an eye on who controls this infrastructure and what it costs. Today's experimental satellite could become tomorrow's competitive advantage, and you don't need to be an astronaut to understand where the technology is heading.
Google is trying to feed its AI from orbit, but the data it processes has to come from somewhere on the ground. Much of that information is already sitting with your phone company, in records far older than most people realize.
US: Your Mobile Carrier Stores Your Call Logs for Years
Your mobile phone company may be keeping records of your calls, texts, and location for years, even after you've forgotten they happened. We're not talking about the content of your conversations. We're talking about who you contacted, when, how long you talked, and which cell towers your phone connected to. Put enough of that together, and someone can reconstruct your relationships, daily routines, workplace, and even visits to doctors or lawyers.
This isn't hypothetical. In 2024, a breach at AT&T exposed communication records belonging to 109 million customers, representing nearly every cell phone subscriber on the network.
The information included phone numbers, call and text activity, durations, and location data dating back to 2022 and 2023. No conversation recordings were involved, but that hardly makes the information harmless. Metadata can tell an incredibly revealing story without anyone hearing a single word you said.
The bigger problem is how long carriers retain this information. In response to an FCC inquiry in 2022, major carriers disclosed that they kept certain geolocation records for as long as five years. These records can be valuable to advertisers, data brokers, and law enforcement.
The more information companies accumulate, the more attractive they become to hackers. The Chinese-linked Salt Typhoon campaign reportedly compromised all major US telecommunications carriers. Privacy experts note that five years of this metadata is detailed enough to reconstruct visits to specialized medical clinics, bankruptcy attorneys, or political events, without recording a single second of audio.
So what's the upshot for you?
Check your mobile carrier's data retention and privacy policies. Don't assume that deleting your phone history deletes the records held by your carrier.
If your communications are sensitive, consider a provider that minimizes how much information it retains. The best way to prevent your personal data from being stolen five years from now is to make sure it doesn't exist five years from now.
As an example in the U.S., one smaller carrier, Cape, says it deletes most call records within days rather than years. Cape hired security firm Trail of Bits for an independent public audit to test whether it actually purges call records within days, as it claims.
The auditors found no evidence the company was violating its stated deletion policies. Records are still held by network partners, but Cape says it limits their ability to connect those records back to individual subscribers.
Your phone company has been feeding on your records for years, and your car has been quietly collecting its own. Your daily drive may already be on file somewhere, and it may be sold to people who set your insurance rates.
US: GM Pays $220 Million for Selling Your Driving Habits
General Motors has agreed to pay $220 million to settle allegations brought by 36 states that it collected and sold drivers' private information without proper consent. The data included precise location information, driving habits, speeding, braking, and even seatbelt use.
GM collected much of it through Smart Driver, a free feature in its connected-car and OnStar apps. The information was then passed to data brokers, including LexisNexis and Verisk, which reportedly sold it to insurance companies.
Enrollment was also shaped by incentives. GM rewarded dealership sales staff with bonuses for OnStar enrollments, and many sales representatives reportedly accepted Smart Driver terms on customers' behalf during car delivery.
More than 370,000 owners were shown an OnStar app screen saying they were not enrolled, even as GM was collecting and selling their location and braking data. GM reportedly earned more than $20 million directly from selling the data to brokers.
Your driving data could potentially influence what you pay for insurance. You buy a car, use the manufacturer's app, and unknowingly create a driving profile that could follow you into your next insurance renewal. GM discontinued Smart Driver in 2024, but the legal consequences have continued.
Under the settlement, GM must obtain explicit consent before collecting or sharing certain driving and location information. It must also have an independent third party assess its privacy compliance.
The company says it will no longer sell vehicle information that can be linked to an individual, their vehicle location, or driving behavior. An earlier Federal Trade Commission agreement also restricted GM and OnStar from sharing certain customer data.
The timing is notable as GM moves toward adding AI assistants, cameras, voice controls, and more connected services to its vehicles. These features promise convenience, safety, and personalization, but they also create more opportunities to collect information about drivers and passengers. Opting out of data sharing may also mean losing access to some features.
So what's the upshot for you?
Check your manufacturer's privacy settings, review what your connected-car apps are collecting, and be careful about agreeing to features you don't need. Modern vehicles log precise trips, driving behavior, cabin sensors, diagnostics, phone sync, some even record the weight of the driver, and much of that data is shared or sold with limited opt-outs. The next time your car offers to monitor your driving for free, remember that free might just mean somebody else is paying with your data.
Your car feeds its data to brokers who feed it to insurers. When that kind of trust gets broken at scale, the losses don't stay small. Binance learned that the hard way, when the trust was placed in a single employee's laptop.
Global: A Compromised DevOps Laptop Caused the $6.3B Binance Breach
Binance, the world's largest cryptocurrency exchange, has suffered a staggering $6.3 billion theft, and the suspected entry point was an employee's compromised laptop. The FBI has attributed the attack to North Korean hackers known as TraderTraitor, a group associated with previous major cryptocurrency thefts. The attackers reportedly used access from a development engineer's machine to reach an internal service involved in processing withdrawals.
The hackers apparently didn't need wallet passwords or private keys. Instead, they exploited the employee's access to a service container connected to Binance's withdrawal process, allowing unauthorized transactions. The theft, reported at $6.36 billion, was taken in roughly 20 minutes. That would make it the largest cyber heist on record, surpassing the $1.5 billion Bybit hack of 2025, which was also attributed to TraderTraitor.
Rather than leaking private keys, the attackers reportedly used the engineer's access to exploit Ethereum's EIP-7702 delegation feature, letting an external smart contract approve withdrawals directly from Binance's hot wallet. Investigators have not released all the technical details, but security researchers say the attack appears to have exploited the trust placed in automated systems.
Binance froze approximately $3 billion in assets and temporarily suspended some withdrawals while it investigated. The company says customer funds are safe and has committed to covering all losses. It has already replenished approximately 95% of the affected assets, with the remaining 5% involving less-liquid cryptocurrencies that may take longer to replace.
The incident has uncomfortable similarities to the 2025 Bybit breach, where attackers also compromised a developer's computer to manipulate cryptocurrency transactions. The FBI says the stolen Binance assets are being converted into Ethereum and could subsequently be laundered through decentralized finance platforms. The investigation also highlights how sophisticated attackers increasingly target employees and trusted administrative systems rather than simply attacking the network perimeter.
So what's the upshot for you?
Your biggest security vulnerability may not be your firewall, encryption, or expensive security software. It could be one employee's laptop with too much access. Limit administrative privileges, isolate development environments, require independent approval for financial transactions, and monitor automated processes.
Trusting your employees is good management, but verifying their access is good security, especially when one compromised laptop can cost $6.3 billion.
Binance's losses came from trusted systems doing exactly what they were built to do, for the wrong people. The next story is about an AI-powered attacker that doesn't need a trusted employee at all. It goes after the investigators who track criminals.
Global: AI Agents Are Now Breaking Into the Systems That Track Crypto Criminals
Chainalysis, one of the world's leading blockchain investigation companies, has been hacked. The interesting part is how the attackers reportedly did it.
The company says an external attacker used what it calls "offensive agentic techniques" to gain access to its internal systems and tools. In other words, AI agents were involved in an attack against a company whose business is helping law enforcement and financial institutions track down cybercriminals.
The irony is hard to miss: Chainalysis is the firm that law enforcement agencies hire to track down cryptocurrency criminals. This appears to be one of the first documented cases of attackers deploying autonomous AI agents to carry out a multi-step intrusion against a cybersecurity target.
The attackers accessed user IDs, full names, email addresses and job roles belonging to people using the platform. Chainalysis says there is currently no evidence that investigation files, transaction records, passwords, multifactor authentication secrets or session tokens were compromised. The company is still determining exactly which customers and records were affected.
Chainalysis hasn't explained precisely how the attackers used AI, which vulnerabilities they exploited, or how much of the operation was automated. That's an important distinction. Agentic AI can carry out multiple steps independently, including identifying weaknesses and executing actions, rather than simply suggesting what a human attacker should do.
The company has brought in outside cybersecurity experts, is patching vulnerabilities and is using AI itself to review and strengthen its defenses. Customers have been warned to watch for unexpected password-reset requests and authentication prompts, and to enable two-factor authentication. The company also confirmed that this incident has no connection to the recent Binance cryptocurrency theft.
So what's the upshot for you?
We're moving beyond AI that helps criminals write better phishing emails to AI that can actively participate in breaking into systems. Make sure your AI tools have tightly controlled permissions, monitored activity and proper authentication. You know, the same old, same old!
AI agents are now hunting the hunters, and they're finding the systems that keep their own records. But the more ordinary threat is the one that comes to the front door, the kind that can't be stopped with a cleverer algorithm. In the UK, that threat arrives in 100 days, when the landline goes silent.
GB Landlines to Go Silent in Britain in 100 Days
Britain's old telephone network is heading for retirement, and Openreach says there will be no more extensions. The deadline is January 31, 2027. That does not mean every landline phone goes dead, but anyone still using the old system needs their provider to arrange what comes next.
For many households, the change means plugging the phone into a broadband router instead of a wall socket. Some customers without a fiber connection may get an alternative that lets them keep using their existing phone line. The point is to replace the aging network, not confiscate your handset.
The bigger headache is everything else connected to those lines. Personal care alarms, fire and security alarms, and business payment terminals can also depend on the old network. Age UK says providers must give vulnerable customers practical help, rather than assume an unanswered letter means everything is sorted.
Businesses face another squeeze. Openreach has doubled the monthly fees on legacy copper business lines, a deliberate nudge to push reluctant companies toward digital IP-based lines before the deadline. Late movers could also struggle to book engineers. The switch-off also affects UK red public phone boxes, emergency traffic signaling systems, and drop-pendant personal alarm buttons used by elderly residents.
So what's the upshot for you?
If you are a Brit with a landline, ask your provider what needs switching, check any connected alarms with their suppliers, and get a plan now. Paying extra for a line heading toward retirement is a lousy subscription strategy.
And now to tie it all together....
Mad cow disease spread because a system fed its animals on the remains of their own kind, and nobody stopped to ask what that was doing to the herd. This week's stories show the same pattern across the digital world.
AI-generated fiction keeps circling the same lighthouse keeper, and researchers worry that machine-written output is feeding the next generation of machines with less variety each time. Criminals are feeding stolen data into AI to decide which victims are worth the effort. They're also feeding on stolen AI keys, leaving companies with bills that can exceed $100,000 a day. Google is building AI data centers in orbit, while the data those systems depend on sits with phone carriers that keep years of records. Cars sell driving habits to brokers, and a single laptop with too much access helped drain $6.3 billion from Binance. AI agents are now breaking into the firms that trace criminals, and in Britain, a whole network of old phone lines is going quiet.
The common thread is that the output of one system keeps becoming the input for another, and each loop collects more data, more trust, and more risk along the way. Nobody has to intend harm for the damage to spread. It only takes a system that consumes whatever is nearby and never checks what it's eating.
So the conclusion is practical. Use AI to sharpen your work, but bring something new to the table. Treat your AI keys and your email as the valuable assets they are. Ask how long your carrier and your car keep your records. Verify who has access to the systems that handle your money. And when a tool tells you what you're worth, remember that the question is who else is listening to the answer.
And that brings us to our quote of the week, from Friedrich Nietzsche, in Daybreak (Morgenröthe):
"The snake which cannot cast its skin has to die. As well the minds which are prevented from changing their opinions; they cease to be mind."
This quote fits the theme of Mad Cow. A system that cannot shed its old patterns keeps repeating them, and the repetition is what kills it.
The lighthouse keeper who keeps showing up in every AI story is a small example. The larger one is a technology industry that keeps training on its own output and calling the result progress, while the people it feeds pay the cost.
That's it for this week; stay safe, stay secure, be nice to cows, and we'll see you in se7en.
Comments
Post a Comment