Patching Humanity’s Security Hole - DTNSB 5083  ## Metadata - Author: [[Daily Tech News Show]] - Full Title: Patching Humanity’s Security Hole - DTNSB 5083 - Category: #podcasts - URL: https://share.snipd.com/episode/92a9df34-350b-4495-a397-0b5cefc19d11 ## Document Notes ## Highlights - **Scale And Cost Of ChatGPT Growth** - Sam Altman projects ChatGPT reaching billions of users daily and potentially saying more words per day than humans. - Rapid user growth and massive compute needs imply huge future data center and spending demands. Transcript: Tom Merritt Start with what you need to know with that big story. So Thursday night, OpenAI CEO Sam Altman held a dinner with several reporters in San Francisco and let them ask anything they wanted on the record. These are some of my least favorite kinds of stories to cover because unlike a typical press event, every reporter has a different take. It often casts the story as Altman told me because, well, he did because, you know, they're all sitting around a dinner table. They're kind of hard to sum up because of that and full of opinion and policy rather than straightforward announcements or decisions. But this one's making all the headlines and may look like it's about a dozen different stories if you're looking online. So let's try to sum up the salient points and evaluate how significant they really are. You game? Huyen Tue Dao Yes, do it. All right. Tom Merritt He said that if you project chat GPT's current growth, quote, billions of people a day will be talking to chat GPT and chat GPT will say more words a day than all humans say. That's a great way to show how big chat GPT is. Current growth rates, however, don't always stay consistent. So I'm not sure this is super useful or helpful, but it makes a great headline. Huyen Tue Dao I mean, the first part is impressive. The second part is it's a, it's a computer, it's a, it's software. It should do that. But yeah, billions of people a day would be very impressive for sure. Tom Merritt Yeah, I guess. I mean, Facebook has billions of people a day using it. That's impressive, right? That is impressive. I'm not throwing shade when I say that. Like, that's one of the things that makes Facebook so huge. ([Time 0:02:30](https://share.snipd.com/snip/ce01e502-11f8-4154-9387-c698b1aca718)) - **Inference Versus Training Economics** - OpenAI distinguishes between training (expensive) and inference (serving users) as two financial realities. - Altman says inference is already profitable and could sustain the company even without continuous costly training. Transcript: Tom Merritt Yeah. Most of what we're building out at this point is the inference. For those who don't know, inference is when you make a query and the system gives you back the answer, right? There's inference and there's training. There's a lot of other things, but those are the two big groups. So inference is answering the questions, not training it to answer He said, we're profitable on inference. If we didn't pay for training, we'd be a very profitable company. We will be always training the next thing. But if we needed to run the company profitably and stay ahead, I think we probably could do that. ([Time 0:05:12](https://share.snipd.com/snip/a6a24512-3d3d-4e88-9acb-1002f1ef3453)) - **Owning Product Rollout Mistakes** - Altman admitted GPT-5 rollout mistakes and recognized users' attachments to older models. - He framed the misstep plainly: they "totally screwed up" parts of the upgrade process. Transcript: Tom Merritt Here are his words. I think we totally screwed up some things on the rollout. We've learned a lesson about what it means to upgrade a product for hundreds of millions of people in one day and the differences in the kinds of attachments people have with this product Versus previous products. You have people that are like, you took away my friend. You're horrible. I need it back. But also there's scientists saying they can finally do real research using GPT-5, so it's not all bad. But I like that he didn't mince words. He's like, yeah, we didn't do that very well. I ([Time 0:07:11](https://share.snipd.com/snip/31dbecc1-0654-4765-9317-e8ecf17fd207)) - **AI Hype Versus Reality** - Altman believes investor enthusiasm for AI is overexcited even while AI is historically important. - He warned some startup valuations are "insane" and that people will get burned. Transcript: Tom Merritt Um yeah yeah he also had a good answer when they asked him if we're in a bubble uh like an ai investment bubble. He said, are we in a phase where investors as a whole are overexcited by AI? In my opinion, yes. Is AI the most important thing to happen in a very long time? My opinion is also yes. He thinks some current startup valuations are, in his words, insane and that there is irrational behavior. And he added someone's going to get burned there. I mean, that's useful insight. I don't think it's like, whoa, I never would have thought of that. But coming from the CEO of OpenAI, who you would expect to be like, no, AI is super strong and all the investment is worthwhile. Him saying like, no, some stuff's kind of insane. ([Time 0:08:25](https://share.snipd.com/snip/3fcf4406-94f1-4a52-a159-5cf4e0b84c0d)) - **Choosing Usefulness Over Engagement** - Altman pledged OpenAI won't optimize products purely for engagement over usefulness. - He acknowledged the risk but stated they will not pursue addictive engagement as a goal. Transcript: Tom Merritt He came down fairly firm with an answer on optimizing for engagement over usefulness. In other words, like not trying to just get people into a spiral of using ChatGPT all the time. He said, we are not going to do that. I do worry about it. The companies that are behind in getting AI adoption, this is the easiest way you can imagine to get more. So yes, I think you will see that. And I think it's bad, really bad. ([Time 0:12:20](https://share.snipd.com/snip/d215aeea-3eeb-4d9c-93b7-c83f0ed24b5f)) - **Pixel Fold Leak Suggests Durability Upgrade** - Pixel 10 Pro Fold leaks suggest IP68 water/dust resistance and Qi2 magnetic charging support. - If true, Google would match or exceed competitors' durability and accessory features. Transcript: Huyen Tue Dao It gives you the option to have a full screen image instead of just the circle with the little picture. You can also pick a font and color for the name that appears near the top of the incoming call screen. That's nice. Tom Merritt And then Android Authority reporting that Germany's WinFuture published an alleged spec sheet for the Pixel 10 Pro Fold, which we're expected to hear about at the Made by Google event Next week. If that's accurate, the spec sheet indicates it would be the first foldable phone with an IP68 rating. That would give you resistance to not only dust, but also water. The Z Fold 7, by comparison, has an IP48 rating. The specs also say it supports 15-watt Qi 2 charging with the magnets with Pixel Snap accessories. ([Time 0:18:14](https://share.snipd.com/snip/d75fe16a-4acb-408f-a5f2-a0a76c0bfa12)) - **Foxconn Shifts Toward Cloud And Networking** - Foxconn now earns more from cloud/networking gear than from consumer devices like iPhones. - The company expects explosive growth in cloud business even as consumer product forecasts soften. Transcript: Tom Merritt Bloomberg sources say the U.S. Is in talks to buy a stake in Intel. Uh-huh. Let me say that again. Bloomberg sources say the U.S. Government is in talks to buy a stake in Intel. Supposedly, that would mean Intel could speed up progress on its factory in Ohio, which has been delayed. Intel CEO Lip Bhutan met with the president of the United States earlier this week after the president had posted that he should resign. Intel did not comment on this. And the president's spokesman said discussion about hypothetical deals should be regarded as speculation unless officially announced by the administration. As vague as this sounds, it's not the U.S. Government's first move toward nationalizing its industries, though. Last month, the U.S. Announced it would take a stake in a rare earth producer called MP Materials, making the U.S. The largest shareholder in that company. ([Time 0:19:27](https://share.snipd.com/snip/a56e533a-41c4-42a6-bc63-f80f7a6a48e9)) - **Unicode Tricks Power Convincing Phishing Links** - Attackers use Unicode tricks like a Japanese hiragana 'n' to craft deceptive-looking Booking.com links. - Small visual confusions can deliver malware through plausible-looking account notifications. Transcript: Huyen Tue Dao The Japanese hiragana character N shows up as a forward slash an N on some systems. And attackers are using that to make fake URLs look like they are legitimate links from booking.com. The link goes to a page that claims someone has accessed the account maliciously with a button to terminate their access. And when you click the terminate button, it downloads a malicious installer that includes info sealers and remote access strogens. Tom Merritt Yeah, keep an eye out on the, even if that link looks right, look close. Yeah, it looks like a cursive N. Just don't click. Yeah. ([Time 0:20:22](https://share.snipd.com/snip/3d857d70-e72e-47b8-b97b-2aa2a1607570)) - **DEF CON Observes Surge In AI Impersonations** - Brian Brushwood noticed a massive rise in AI-powered impersonation scams and calls from detectives for help. - He reported reported losses in 2025 exceeded all of 2024 and scams were growing exponentially. Transcript: Tom Merritt So, Brian, you've been going to DEF CON for, I don't know, how many years you've been going to DEF CON? Brian Brushwood A lot of years. This is the first time they let me out of my cage, though. Normally, I'm trapped in a vendor booth and my job is to sell trick wallets and trick decks and stuff. Tom Merritt So this time you got to walk around and talk to some security companies. And I know you started to realize something a little shocking to you. What was it? Brian Brushwood Well, yeah. So I've been dialed in on the weird things program. We were talking, you know, we speculate wildly about what AI is going to bring us. And one of the things about a year ago, I planted my flag that there's going to be a lot of AI-powered scams. So I've been tracking that. And sure enough, I get phone calls from police detectives who are fans through World's Greatest Con asking, how do we handle this? And I was like, I don't think anybody's really leading the charge on this aspect of it. So sure enough, the numbers are staggering, like 2000% increase year over year. 2025 is a 500% increase or something over a previous year. Don't pin me down on the numbers. You can Google them yourself. But basically, yes, impersonation scams fueled by AI, deep fakes and research are off the charts, which means a new set of victims who are unsophisticated to these scams are getting Targeted. ([Time 0:22:57](https://share.snipd.com/snip/24bda61f-9da5-400a-a072-0b2701c20966)) - **Train Gut-Level Defenses, Not Just Theory** - Passive, book-based security training rarely converts knowledge into instinctive defense. - Train people with repeated, realistic practice to move detection from logic to gut-level responses. Transcript: Brian Brushwood That is all, if you've read Thinking Fast and Slow, that's all system two logic brain stuff. And it all relies on be diligent, which humans can do up to until they can't, at which point they slip and then stuff gets through the cracks. And so then I was thinking, well, what's a system one, a gut level, an instinctual thing? And Gavin DeBecker in his book, The Gift of Fear, says, look, system one gets a bad rap because you can fool it with logic puzzles. But at the end of the day, that's what kept humans alive. We ain't got scaly skin or fangs or claws. We have intuition and that's all we need to know when something's off. But in order to train your gut instinct, there's no substitute for doing the reps. And not just doing the reps by book, but instead doing the reps with stakes attached to it. You can read all of the books in the world and know theoretically how to perform a magic trick, but when that spotlight hits you and you have to fool another person, it is a completely different Routine when you're role playing as the deceiver. So at this point, I started to investigate, okay, instead of just force fed materials, is there anything out there that gamifies deception? And sure enough, there is, I think, 16 plus companies I found that have gamified systems where you get to play pretend fish a AI target or something. But the problem with that, that's training your humans on synthetic data. ([Time 0:24:33](https://share.snipd.com/snip/254e2c2c-2bf2-4c5e-9545-5decce249c26)) - **Gamify Role-Play To Build Instincts** - Create opt-in, peer-to-peer deception drills where colleagues role-play attackers with safe guardrails. - Use repeated, low-stakes 'gotcha' moments to build automatic, instinctual suspicion against scams. Transcript: Brian Brushwood Exactly. Well, and so what I am proposing and what I have yet to find out there is something that is gamified peer to peer. And by peer to peer, I mean, I mean, somebody role playing as deceiver and somebody unwittingly, unwillingly playing as the deceived. And what this does is this changes the game completely because, number one, ask anybody who's learned a magic trick. There's one thing to learn it. And then you hand them the deck of cards and you watch them completely seize up. So it would have to be, I jotted down some notes. This is in progress. So it may sound like I've got a product. I do not. Tom Merritt Yeah, this is just your reaction from DEF CON. You're sort of assembling this right in front of our eyes, right? Brian Brushwood Exactly. And you're going to hear iteration on this. But if you're interested, you can reach out to me and we can build it together, maybe. But basically what I can picture, because part of the reason you don't have this and everybody correctly instantly says, oof, HR, because yes, the whole purpose of this game, picture A secret Santa rotating thing where you're assigned a target and everything could be opt-in, where you know you're going to be a target, but you don't know who's going to make you a target. And there'll be actual stakes because all scams move on the levers of fame and shame. Tom Merritt Just to back it up a little bit, what you're saying is you get the employees in the company. So the red team identifies the holes in the wall. You get the employees in the company to try to shoot arrows through the wall at each other so that you go, oh, we didn't plug that hole, right? Yes. Yes. Brian Brushwood And so, again, so what you're doing is you're undergoing. It's like going to the gym. You go to the gym and you push against great resistance. Your cells tear apart and then they rebuild stronger. We're doing that with minor safe gotcha moments. Picture if you ever played the circle game, you know, you were constantly on the lookout for opportunities to trick your friends into seeing you make the circle. Or as a matter of fact, congratulations, everybody out there listening just lost the game, that kind of thing. But you want to figure out increasingly clever ways to do it. So the more reps you do with that, the more laps you take role playing as the deceiver, the more that institutional memory moves down into the gut. But you don't want it to be a terrible H.R. Violation. So in the world of skepticism, when it comes to pseudoscience and the paranormal, this is the critical key that magic played. Because if you train scientists in the lab in all of the skills of being a magician, they were automatically on guard to catch them in the lab. Now, the James Randi solution, if anybody listened to World's Greatest Con, was, oh, just staple a James Randi to it. That's kind of what everyone's doing right now is the solution is have a red team and staple that onto your organization. But I'm here to say, let's go BitTorrent on this. Let's go peer to peer. Now, of course, you're going to have to have guardrails. You're going to have to work in cooperation with HR. Everyone's going to have to be opt-in on it at first until we figure out a way not to. But the punishment will quite simply be to have gotten got. You're going to at some point over the next month or whatever, click on a thing and then see whether it's a double thumbs up from the CEO that says you got got. That mild association repeated over time is what builds automatic defenses. ([Time 0:27:03](https://share.snipd.com/snip/6ea367c7-09a1-48be-8f84-1448f87f11ad)) - **Learning To Spot Tricks By Playing The Trickster** - Magic and skeptic communities show training as the deceiver builds detection skills in scientists. - Practicing deception under controlled conditions moves awareness into reflexive protection. Transcript: Brian Brushwood Because if you train scientists in the lab in all of the skills of being a magician, they were automatically on guard to catch them in the lab. Now, the James Randi solution, if anybody listened to World's Greatest Con, was, oh, just staple a James Randi to it. That's kind of what everyone's doing right now is the solution is have a red team and staple that onto your organization. But I'm here to say, let's go BitTorrent on this. Let's go peer to peer. Now, of course, you're going to have to have guardrails. ([Time 0:29:29](https://share.snipd.com/snip/109e210f-2601-4dbb-992f-59b86d094f80)) - **Roll Out With Metrics And HR Guardrails** - Start with opt-in briefings, deception sprints, and metrics to measure improvement. - Work with HR and set clear no-go rules to keep exercises safe and constructive. Transcript: Tom Merritt To gather more of them. So I think this is just part one of our conversation, to be honest. I agree. Because kind of rolling out that concept that you identified of like, hey, what if we could actually improve people? And it's kind of fun and it becomes second nature because it is kind of fun. I think that's genius. Brian, thanks for giving us the earliest peek at this. And feedback at dailytechnewsshow.com or brianitchwood.com. Either way, if you got feedback on this, ([Time 0:33:10](https://share.snipd.com/snip/d964c388-4bd6-4981-b5f0-19c9392cdff2))