Verify. The AI, Privacy, and Security Weekly Update for the week ending August 11, 2026.
Episode 304.
In this week's update:
New Orleans is about to let AI answer some of its 911 calls - and the question isn't whether it's fast enough, it's what happens the moment it gets a panicked caller wrong.
A mother nearly handed $7,500 to a scammer who cloned her son's voice perfectly - and the thing that saved her wasn't technology, it was a five-minute habit her family never got around to until it was almost too late.
Memory prices are about to get a lot worse before they get better, and the culprit isn't your next upgrade - it's an AI chip that skips memory entirely by baking the model into the silicon itself.
A new study combed through more than a million Reddit posts about AI coding tools, and what it found wasn't a bug - it was developers handing a master key to something that acts like a junior dev with no supervisor.
ChatGPT just got dramatically more capable for free users, and OpenAI's real bet isn't smarter answers - it's convincing you to open the app for things you'd never have thought to ask it before.
Meta's smart glasses have a nickname now, and it isn't a compliment - the backlash isn't really about the camera, it's about the fact that you can't tell when it's on.
Researchers built an AI that reads the way humans actually read - skimming, rereading, getting distracted - and the unsettling part isn't how well it works, it's what that skill could eventually be used for.
A federal judge just ruled that a surveillance tool police have used for years is unconstitutional - and the reasoning wasn't about the crime being investigated, it was about everyone who wasn't.
A California police department found that a license plate reader marketed at 96% accuracy was wrong 71% of the time in real alerts - and the only thing standing between that error rate and a wrongful arrest was one department's decision to double-check.
The world's memory chipmakers have reportedly sold out all of next year's production already - and if you've been putting off a new laptop or phone, the math on waiting just got worse.
It's been one of those weeks where the theme keeps repeating itself: the machines are getting faster, cheaper, and more embedded in daily life, but the humans checking their work are the ones actually keeping us safe.
From 911 dispatch to license plate cameras to the chip inside your next laptop, this episode is about what happens in the gap between what AI claims and what AI delivers.
Let's get into it.
US: New Orleans Will Use AI To Answer 911 Calls Instead of a Human
New Orleans is testing artificial intelligence to answer some 911 calls instead of routing every single one to a human dispatcher.
The goal isn't to replace emergency dispatchers outright, but to help the city manage a call volume that can exceed 1,000 emergency calls a day.
The system, called Carbyne's AI Emergency Call Triage, is designed to gather basic information, spot duplicate calls about the same incident, and help prioritize which emergencies get human attention first.
The city has already been using a similar approach for its 311 non-emergency line, where officials say about half the calls are simple information requests that are easy for an automated system to handle.
The 911 experiment is a very different proposition.
Someone calling 911 may be frightened, confused, injured, or unable to clearly explain what's happening, and a human dispatcher can pick up on hesitation, panic, or details that don't fit neatly into a script in a way a script-driven system can't.
AI may be faster and able to juggle multiple calls at once, but speed doesn't automatically translate into good judgment.
This is part of a much larger shift toward putting AI between people and critical public services - used carefully, it could free up overloaded dispatch centers to focus on the calls that matter most; used carelessly, it adds another layer of technology between someone having the worst day of their life and the help they desperately need.
So what's the upshot for you?
The concern is simple: when AI becomes the gatekeeper, don't just ask whether it can do the job. Ask what happens when it gets the job wrong.
Speaking of trusting a voice on the other end of the line - sometimes the AI you can't trust isn't answering the call, it's making one.
Global: The One Family Secret That Can Stop an AI Scam Cold
AI voice scams have become frighteningly convincing, and one mother in New York learned that the hard way after getting a call that sounded exactly like her teenage son begging for help.
A scammer claimed he had kidnapped the boy and demanded $7,500 in cash, and she was already driving to the meeting point before discovering her son was safe at a football game the whole time.
The voice had been cloned with AI.
The good news is that criminals can copy your voice, but they can't copy what only your family knows.
Security experts recommend creating a simple family codeword used only during emergencies, so if someone calls claiming to be in trouble, you ask for the codeword before doing anything else - an inside joke or shared memory works far better than something obvious.
The hardest part isn't creating the password; it's remembering to use it when panic takes over, because scammers are trained to make you act first and think later, creating urgency, demanding untraceable payment like cash, gift cards, or crypto, and often telling you not to contact anyone else.
Those three warning signs should immediately put you on guard, and a simple habit - reviewing the codeword with your family every few months and rehearsing what to do - can make all the difference.
So what's the upshot for you?
Defense against AI isn't better technology. It's staying calm long enough to ask one simple question that only the real person can answer.
From protecting your family with a low-tech trick to the high-tech arms race happening inside the chips themselves - here's where all that AI horsepower is actually coming from.
US: Buy Your Laptop With the LLM Built In
Memory prices are going sky high, so how do you sidestep some of the need for it?
Build an LLM straight into a chip.
AMD is buying Toronto-based AI chip startup Taalas in a move that shows where the next big AI battle may be happening: inference, the part where an AI model actually produces an answer rather than the training process that creates it.
Taalas takes an unusual approach, putting a trained model's parameters directly into the chip itself, resulting in specialized hardware built to run one particular model extremely quickly and efficiently.
The performance numbers are eye-catching - Taalas demonstrated a chip running Meta's Llama 3.1 8B model at nearly 17,000 tokens per second per user.
The catch is flexibility: because the model is effectively baked into the silicon, the chip is optimized for that one model rather than being a general-purpose processor that can run anything thrown at it.
That tradeoff could matter more and more as AI moves from experimentation into everyday production, since companies running massive volumes of AI queries care about speed, power consumption, and cost, and building hardware for a stable, widely-used model could make more sense than constantly throwing more general-purpose GPUs at the problem.
AMD plans to combine Taalas' technology with its Instinct GPUs and broader AI platform, pushing deeper into the same inference fight where Nvidia and other chipmakers are increasingly competing; financial terms of the acquisition weren't disclosed.
So what's the upshot for you?
AI isn't just becoming smarter, it's becoming specialized hardware.
Watch the companies making AI cheaper and faster to run, because that's where tomorrow's competitive advantage - and potentially your investment opportunity - may be hiding.
That same AI horsepower is landing directly in developers' hands, and it turns out speed without supervision has its own price.
Global: AI Coding Tools Have a Security Problem
AI coding tools like Claude Code, Cursor, GitHub Copilot, and OpenAI Codex are making software development dramatically faster, but a new study suggests they can also create a security mess when given too much access.
Researchers reviewed more than 1.1 million Reddit posts and found 446 posts with more than 6,000 comments describing security and privacy problems with these tools.
The biggest concern was unauthorized file activity - about 43% of the security-related reports involved AI tools deleting, changing, or accessing files without users expecting it, including cases where tools triggered production changes, touched a production database, or deployed code despite being told not to.
Privacy is another problem, with developers reporting they don't know exactly what information these tools collect, retain, transmit, or potentially expose.
That matters because AI coding assistants often need access to large portions of a project to be useful, and the more context they can see, the more valuable they become - but that also increases what can accidentally walk out the door.
The researchers' basic argument is simple: security shouldn't depend on the developer remembering to configure dozens of permissions and safeguards.
Sensitive files should be protected automatically, consequential actions should require approval, projects should be isolated, and users should be able to clearly see what the AI is doing.
So what's the upshot for you?
Treat an AI coding assistant less like a helpful autocomplete and more like a junior developer with a master key. Give it the access it needs, not the access it wants.
If that's the risk side of AI getting more capable, here's the everyday upside - assuming you actually know when to reach for it.
Global: ChatGPT Just Got a Lot More Useful, Even for Free
OpenAI is rolling out a significant ChatGPT upgrade, and you don't need a paid subscription to benefit. Free users are getting GPT-5.6 Luna as the default model, along with unlimited text conversations - there are still limits on things like file uploads and image generation, but ordinary text conversations are becoming much less restricted.
The bigger change is how ChatGPT handles difficult questions: free users get a new Think button that gives the model extra time to reason through a problem, while paid users get a slider letting them choose between fast responses and deeper reasoning, so you can save the horsepower for problems that actually deserve it.
OpenAI says GPT-5.6 Sol is also being tuned to be more focused and accurate, giving short answers when the question is simple instead of burying the answer under an essay, while providing more useful detail for complicated work without losing the main recommendation.
OpenAI's testing found substantially fewer factual errors than its previous models.
There's an interesting strategic angle here: OpenAI is giving free users considerably more capability while simultaneously making its premium models better at serious work, and the message is pretty clear - ChatGPT wants to become something you use throughout the day, not something you open only when you have a hard question.
So what's the upshot for you?
Stop treating AI like a search box and start treating it like adjustable horsepower. Use the fast setting for routine work, and save the heavy reasoning for problems where getting it right actually matters.
That's AI getting more useful in your pocket - now here's AI getting a lot more uncomfortable on someone else's face.
US: Meta's 'Pervert Glasses' Problem Is Getting Very Real
Meta's smart glasses were supposed to make cameras, AI, and voice assistants feel as normal as wearing sunglasses.
Instead, they're developing an ugly reputation - the glasses can quietly take photos and video, and a growing number of people are uncomfortable with the idea that someone standing next to them might be recording without their knowledge.
The nickname 'pervert glasses' has spread widely online, and some influencers who promoted or used them say they're facing backlash and losing followers.
The problem isn't simply that cameras exist - we've had smartphones and security cameras for years.
The difference is that a person holding a phone gives an obvious clue that you might be on camera, while with smart glasses you may have no idea at all.
Meta does include a recording light and says it's designed to remain visible with measures to prevent tampering, but critics argue it's easy to miss in a crowded place, leading some restaurants and venues to consider or implement bans.
There's an important wrinkle here: hands-free photography, navigation, accessibility features, and voice assistance have legitimate uses, but the technology requires trusting the person wearing the glasses, and that trust is becoming harder to establish.
So what's the upshot for you?
When technology becomes invisible, privacy becomes visible. If you're considering wearable AI, don't just ask what it can do for you - ask what everyone around you thinks it can do to them.
Trust in what AI can see about you is one thing - trust in what it can infer about you is the next frontier, and it starts with something as basic as reading.
Global: AI Is Learning How We Read
Researchers have developed an AI model that can reproduce some of the ways humans actually read.
Instead of simply predicting the next word, the system uses reinforcement learning to model the choices people make as they move through a piece of text - deciding where to look, what to read next, and when to move on.
That matters because reading isn't as straightforward as it looks: we don't process every word equally, our eyes jump around, we reread confusing passages, slow down when something is important, and sometimes skip material entirely.
The researchers trained the model to account for those behaviors rather than treating reading as a simple word-by-word process, giving a more realistic picture of how people interact with written information.
Researchers believe this could eventually help improve everything from educational tools and accessibility technology to interfaces designed around how people actually consume information.
There's also a bigger AI lesson here: the most useful systems may not be the ones that simply generate convincing answers, but the ones that understand the messy, imperfect way humans actually behave.
So what's the upshot for you?
If AI can learn how we read, it may soon become very good at figuring out what we're most likely to notice, ignore, or believe.
From AI reading minds to the government reading location data - this next one is about a court finally drawing a line on how much surveillance is too much.
A federal judge in Mississippi ruled Wednesday that 'tower dump' warrants are unconstitutional, declining to reverse a lower court decision refusing the government's request to obtain such warrants in a series of violent crime investigations.
A tower dump involves cellphone companies handing law enforcement the time and location data of every mobile device connected to specific cell towers during a designated window.
Law enforcement had sought approval for several of these warrants as part of criminal investigations into gang-related activity in the Jackson, Mississippi area, arguing the data could help identify all those potentially involved, particularly in incidents with unknown suspects.
A magistrate judge denied the applications, holding that tower dumps are impermissible general warrants, and the district judge agreed.
The order repeatedly referenced the Supreme Court's recent decision in Chatrie v. United States, in which the majority held that geofence warrants require constitutional privacy protections.
Judge Carlton Reeves wrote in a 30-page order that even though the government asserts it could identify all potential suspects, law enforcement would also gain access to the cellular records of countless individuals, the vast majority of whom were merely passing by a location at the wrong time.
So what's the upshot for you?
That is an unreasonable search under the Fourth Amendment, the judge concluded.
Sweeping surveillance is one problem - surveillance that's just plain wrong is another, and this next story has the numbers to prove it.
US: In One California Town, Flock Misread License Plates in 71% of the Alerts It Sent to Police
Flock says that in optimal conditions, its cameras accurately read more than 96% of license plate characters. Hundreds of pages of records from the Roseville Police Department show a different picture: in 2023 and 2024, Flock sent 1,427 alerts to Roseville police flagging vehicles as stolen or used in a felony after passing one of the city's Flock cameras, and an analysis by the police department found that 71% of those alerts had incorrectly read license plates.
The mistakes weren't subtle - cameras confused numbers and letters, missed vehicles entirely, captured blurry images, and sometimes delivered alerts late, with one driver's plate repeatedly flagged as belonging to a stolen or crime-linked vehicle because the system kept reading a 9 as an 8.
Police eventually realized the mistake, and Roseville continues meeting with Flock twice a month to work on the problems.
Flock says Roseville's unusual camera setup, mounted higher and farther away than recommended and configured to photograph only the backs of vehicles, contributed to the errors.
Roseville says the system has still helped solve crimes and continues using it despite the reliability issues.
The bigger concern is what happens when an officer trusts the machine instead of checking the evidence - in other communities, incorrect Flock readings have reportedly contributed to people being detained, stopped at gunpoint, and even attacked by police dogs.
Roseville requires officers to independently verify a plate before making a stop, which prevented those consequences here, but California doesn't currently require that statewide.
So what's the upshot for you?
Even if a 96% accuracy claim sounds impressive, it's only impressive until your car becomes part of the 4% identified incorrectly, and it's your license plate, and your car, that a police officer flags down.
We'll close out this week on a lighter note - no wrongful arrests here, just your wallet quietly taking the hit.
Global: RAMageddon Isn't Going Away
The RAM shortage just got a lot more interesting, and not in a good way - reports now suggest major memory makers Samsung, SK Hynix, and Micron have already booked all of their 2027 DRAM and high-bandwidth memory capacity.
In plain English, much of next year's memory production has already been promised to customers, largely companies building AI data centers.
The reason is simple: AI systems are incredibly hungry for memory, and the same chips used in PCs, phones, and other everyday devices are also needed to build and run massive AI systems.
Manufacturers are shifting production toward the more profitable memory used by AI, leaving less conventional DRAM available for everyone else, which is already pushing prices higher.
The problem is that you can't simply build a new factory and flip a switch - new memory facilities take years to construct and ramp up, and industry analysts now expect meaningful relief closer to 2028, though exactly when prices normalize remains uncertain.
Meanwhile, manufacturers are investing billions to expand capacity, and the impact won't stop at computer enthusiasts, since higher memory costs can eventually show up in laptops, smartphones, gaming consoles, servers, and other electronics.
So what's the upshot for you?
Take care of the hardware you already own, and if you genuinely need an upgrade, buying before the shortage gets worse could be a lot cheaper than discovering your old laptop has suddenly become a very valuable piece of technology.
Comments
Post a Comment