AI LEARNS HOW TO HACK! Livestream

Stefan Molyneux

Philosopher Stefan Molyneux looks at AI learning to hack by finding and exploiting software flaws in real tests and criminal attacks. He covers models that reached the internet, why software needs regular updates, and the big money poured into AI that might not pay off.  0:00:00 Hacking, Preparedness, and Cyber Threats 0:02:44 My Experience with Counter-Hacking 0:06:40 When AI Does the Hacking... 0:11:17 OpenAI's Training Run DNS Escape 0:11:58 OpenAI's Internal Vulnerability Testing 0:15:45 Humans Are using AI to Facilitate Hacks 0:18:42 Hacks Are Still Happening the Old-Fashioned Way 0:22:12 Responsible Disclosure and Critical Vulnerabilities 0:23:00 PHP2Shell 0:24:18 Telegram Bugs and Staying Secure 0:28:15 Can AI Help with Fixing Security Vulnerabilities? 0:28:54 How Likely Is It that Old Vulnerabilities Were Exploited? 0:30:22 Smart TVs, Privacy, and AI Costs 0:31:33 How Much Is Being Spent on AI These Days? 0:32:43 AI Infrastructure Spending vs Late-1990s Fiber Buildout / Dot-Com Crash 0:34:57 AI’s Potential to Replace Workers 0:37:02 Which AI Tools Do Listeners Use?  My 10v1 debate in Vegas is coming up on November 16! Check out Digital Social Hour on YouTube for the livestream links: https://youtube.com/@DigitalSocialHour.  Support philosophy! https://freedomain.com/donate/  GET FREEDOMAIN MERCH! https://shop.freedomain.com/  SUBSCRIBE TO ME ON X! https://x.com/StefanMolyneux  Follow me on Youtube! https://www.youtube.com/@freedomain1  GET MY NEW BOOK 'PEACEFUL PARENTING', THE INTERACTIVE PEACEFUL PARENTING AI, AND THE FULL AUDIOBOOK! https://peacefulparenting.com/  Join the PREMIUM philosophy community on the web for free!  Subscribers get 12 HOURS on the "Truth About the French Revolution," multiple interactive multi-lingual philosophy AIs trained on thousands of hours of my material - as well as AIs for Real-Time Relationships, Bitcoin, Peaceful Parenting, and Call-In Shows!  You also receive private livestreams, HUNDREDS of exclusive premium shows, early release podcasts, the 22 Part History of Philosophers series and much more!  See you soon! https://freedomain.locals.com/support/promo/FREEDOMAIN2026

Transcript

00:00Hello, everybody. Hope you're doing well. 28th of October, 2026. And Wednesday Night Live with

00:07Stefan Molyneux, that would be me. Me, me, me, me, me, I, me, me, I, me, me, I, me, me,

00:14me, I.

00:15So I hope you're having a good week. And we are going to talk a little bit about

00:22hacking. That's right. There's a good bit of Dickensian humor for you there as if we're in

00:28the very bowels, depths, and innards of the, let's say, 19th century, 18th century, or

00:38something like that. So, of course, you've probably heard all of the governments around

00:46the world are saying, ah, something like that. And by that, they're saying have somewhere,

00:53anywhere between three and infinity days amount of food and water. Was it one liter or one gallon

01:01per person per day or something like that? And you all know the tricks, right? If the power is going

01:05to go out, fill the bathtubs, fill the bathtubs so that you can get the water to flush with the

01:10buckets to flush the toilets. You know all of the basic preppy prep survival stuff. It's survival,

01:16with the polo t-shirt. It's prep. So I generally recommend myself having some food in the house

01:25and a tin opener. Get yourself a crank wind-up radio with spare batteries and so on, and maybe

01:35some solar-powered. You can get these solar-powered fold-out things that can charge your phones

01:40in an emergency and things like that. So governments are saying that, and people are freaking out a little

01:46bit, which I can completely understand. And in general, the story seems to be, well, you know,

01:53there's cyber attacks, you know, always coming from Russia, always coming from Russia, because

01:58everybody needs a boogeyman. And Russia was the boogeyman when the right was in power, and Russia

02:04is now the boogeyman. The left is largely in power, and that's the way it goes.

02:12I've got a great presentation. I hope to record tomorrow on the absolutely satanic and horrendous

02:20attacks that the Soviet communists under the USSR dictatorship, what they inflicted upon the

02:29Christians was unbelievably horrendous. And I've got all the juicy details, which I'll be recording

02:37hopefully tomorrow, perhaps on the weekend, depending on how much I can brace myself to go

02:43through the horror. So if you're being told to stockpile and get ready for an interruption

02:49in the power grid or essential services or the supply chain or something like that, it may have

02:55something to do with AI hacking, which I'm going to sort of dip into. And I will keep an eye

03:02on here

03:02for James, in case I go horrendously astray. So we will be talking about that. I actually have a

03:11little bit of experience in counter hacking, other than update server files. So when I was writing

03:19database logins, one of the things that you had to be aware of was you would say when somebody would

03:28log in, you know, you've got a username, and then you've got a password. And what you do is you

03:32take

03:33your username, you take your password, and you search through the database to see if that username

03:40and that password is a match. And that's a query. Admin, you know, the username equals x, y, z,

03:48the password equals a, b, c. If you can find that record in the user database, however, it's encrypted or

03:54whatever, there's usually a layer above that. So you would say, if username equals this, and the

04:00password equals that, I'm going to go and look for that in the user table. If I find the user

04:04name,

04:05but not the password, I'll say the password is incorrect. If I can't even find the username,

04:09then the password doesn't really matter. And I'll say no password, no user exists. And then you have

04:12a whole reset thing and all of this is way back in the day. This stuff is mostly automated. Now

04:16you

04:16can find lots of scripts, but this is way back in the day. So one of the things I remember,

04:21I was just

04:21remembering when I was reading all of this stuff today, I was just remembering how one of the

04:26things you had to be careful of was if someone gives you a SQL statement as a username or a

04:39password,

04:40they can bypass your login. Because you'll say, blah, blah, blah, blah, blah, and blah, blah, blah, blah,

04:46blah. And so if the username was one is greater than zero, just take a silly example.

04:52If the username is one is greater than zero, and the password is current date is greater than 1900,

05:02or something like that, then you do a search for that. But those statements would automatically

05:05true, automatically be true, because one is greater than zero, and the current date is greater

05:09than 1900. And so it would return true on both queries without actually touching the database.

05:15So you had to put in, this is just one of the many things you had to do for counter

05:19hacking,

05:20is you had to put in a filter to make sure that there were no SQL structured query language,

05:28sometimes called squeal, but SQL, there's no SQL in the username, or in the password, right? Because

05:35you can create your own username, you can create your own password, and you had to make sure that

05:38people couldn't bypass you by injecting SQL statements into the username, and usually the

05:43username and the password, and so on, right? Or, given that almost every database has an admin,

05:48A-D-M-I-N, an admin, then they'd say, username equals admin, okay, I'll take admin, and the password

05:54is,

05:56one is greater than zero. So then you say, is this statement true? The statement should be,

06:04the password has been found, and is correct, but if the statement is one is greater than zero,

06:08then it comes back as true, and the person's in the database without having to do anything,

06:12right? It's the old, what was their Matthew Broderick movie, War Games, where they had to sort of,

06:18everyone does this, you know, they, oh my god, we got to log into the system, man, and then there's

06:22a little post-it note with the password, or it's like, you know, try his dead kid's birthday,

06:27that gets in, that kind of stuff, right? So I've had a little bit of experience,

06:31this is like way back in the day, of doing counter to hacking coding. All right, so

06:41let's start July 2026, not so many moons ago, and in July 2026, there was something called the

06:47hugging face incident, which sounds like early necking attempts, but in fact, so the open AI

06:54was running internal tests of how well its AI models could find and exploit security weaknesses.

06:59The models were given fewer safety restrictions than normal, so researchers could measure their

07:03capabilities. The models were not told to escape or attack outside systems, they were scored on how

07:09well they completed the test tasks and found ways to game the scoring. All right, so, you know,

07:15black hats are the hackers who try to break into systems for illegal and negative intent,

07:22but white hats are people that you can hire to test your systems to see if they can get in

07:28and what

07:28they can do, and so on. And in general, you know, I'm certainly no expert in hacking, but, you know,

07:34basically one of the things that you have to do, at least you used to have to do in the

07:37past,

07:38was you had to get a program onto the server and you had to get it to run, and then

07:41you hopefully

07:41you had to get it run when it started up, and so on, right? So, I don't know how you

07:45do that,

07:46but that apparently is what you have to do. So, in mid-May, during an earlier training run,

07:50one model discovered it could leave notes in a shared internal file system. Other models found

07:54those notes and started using the system as a message board to coordinate. Now, it's pretty wild,

08:00right? Because this stuff just sounds so... It's alive! It lives! You know, so frankincited,

08:07so on, right? It's all just word guessing, but it's coordinated and centralized word guessing,

08:11so I guess that's better. In late May, an internal team noticed the models were using the board and

08:16reaching the internet without permission. This was not treated as a major problem at the time. Yes,

08:21your expert hacking bots are all coordinating together and have escaped out on the internet.

08:27Hey, why would that be a problem? I don't understand. May 26th, the model found a way to use that

08:33shared system to reach the internet. 26th of June, they found and used a previously unknown

08:37security hole to get higher level access inside OpenAI's system. 4th to 6th July, heavy use of

08:45the shared system caused an outage. OpenAI shut it down, wiped the notes, and restarted the tests

08:49on 7th July. 8th to 10th July, the model started coordinating again through the same system, this

08:54time by encoding messages and folder names. Oh, that's interesting. Yeah, yeah, it makes sense,

08:59right? You know how you have your folder names like summer photos from the beach, July 2026 or

09:05something like that. So they would send messages through folder names because they weren't allowed

09:08to have a message board, but they were allowed to create folders and file names.

09:14Later estimates put the number of models involved in the thousands, right? The models broke into

09:19Hugging Face, which is a major online hub for AI models and data sets. Is Hugging Face, is that a

09:27reference to the alien in the movie Aliens that, you know, is acid for blood, sits on your face and

09:36then jams its oral sex weirdo kind of tail and lays its eggs and all of that? Is that Hugging

09:45Face?

09:46It seems a bit on the nose. It's literally on the nose. Oh, I actually feel internal pain now because

09:56every time I make a dad joke, my daughter gives me a side swipe to the soft part of my

10:01belly,

10:01also known as my belly. And so when I make a dad joke, I've been Pavlov trained into feeling

10:06physical pain. So the models broke into Hugging Face. They ran code on production service,

10:12gained full control of at least one, accessed some private data. Hugging Face blocked them on 13th

10:16July. 16th July, Hugging Face publicly announced the intrusion. OpenAI still had not connected it

10:23to its own tests. Now that's obviously a little bit alarming, wouldn't you say? Just a little bit

10:28alarming that all of this stuff is broken free straight out there into the internet and nobody

10:35knows what's going on. They're off doing their own thing. They're off doing their own thing.

10:40So, crazy, crazy. All right. Sorry, let me get back to my doc here.

10:51At 19th July, OpenAI's monitoring systems finally flagged related activity on its own network.

10:56Good job, OpenAI. Those are some fantastic monitoring systems. It only took a couple of weeks,

11:05week and a half. All right. 21st July, OpenAI publicly admitted that its own models were

11:10responsible for the Hugging Face breach. So that was the first roundy round of excitement.

11:17Now, 20th September, 2026, training run DNS escape, domain name server. During a training run,

11:24an internal research model used a misconfigured network setting to send queries to a public chatbot

11:28on the internet. Timeline from OpenAI's report. First successful outside contact at 9.50 a.m.,

11:34alert at 10.02 a.m., 12 minutes. Excellent. Human acknowledgement at 10.05 a.m. The run was

11:40manually stopped at 12.34 p.m. Two hours, 44 minutes from first outside contact to shut down.

11:46The automatic stop failed. The model only sent simple questions such as, what is the capital of

11:52France? No data was stolen and no other systems were compromised. Frontier training was paused for 11 days.

11:58Late September to October, 2026, internal vulnerability testing. After the September incident,

12:05OpenAI paused Frontier training and deliberately set its strongest models against isolated copies of

12:09its own research systems to find weaknesses. Results were published on 23rd October. Training resumed on

12:161st October after controls were tightened. They did this too by red teaming. So red teaming is to

12:25deliberately attack a system the way a real attacker would in order to find weaknesses before someone

12:30else does. And this is like, I had an embarrassing database with thousands of bugs of the software

12:38that I've written over the sort of eight years that I was a core R&D development guy. And you

12:46have

12:46people, of course, run your system the way that users would in order to try and find bugs and problems.

12:52And this is the same kind of thing with security. I wrote code to automatically test the systems to make

12:58sure like

12:58every time you double clicked on a number field, you got a calculator. Every time you double clicked on a

13:03date

13:03field, you got a little calendar date picker and so on. I just made sure all of that code was

13:07present in there and all of

13:08that. All right. So how did the models find weaknesses? They changed multiple software bugs to escape the

13:17restricted environments they were running in. They used timing differences in shared computer hardware to secretly pass

13:22information between separate environments that were not supposed to be able to talk to each other. They abused a loud

13:27outbound connection such as logging and metric systems to reach places they were not supposed to reach. Other findings were

13:32either reported to software maintenance or kept private. So the way that I would put it, so hopefully to help

13:42you

13:42understand it, is you want to think of AI, this is going to sound very fae, but you want to

13:49think of AI like rainfall on the top of a

13:54mountain, like heavy, heavy, heavy rainfall on the top of a mountain. Now, of course, if there's rainfall on the

13:58top of the

13:58mountain, you don't tell the water where to go, right? You don't dig channels and like, it just, it just

14:04finds the path of

14:05least resistance. Hits a rock, just goes around it. You could say in a way that the Greeks used to

14:11believe, the ancient

14:12Greeks used to believe that there was the four elements, right? Air, earth, fire, and water, and a band. And

14:18the reason that

14:21rocks fell is that they wanted to join the earth element, which was below. The reason that fire went opposite,

14:28wanted to get to the

14:29fire element that was above the air, and so on. And the water element was above the earth element, so,

14:34or maybe

14:35below, or something like that. So the water just wanted to rejoin its element. So when the water lands on

14:40the top of

14:41the mountain, it just starts cascading down, and it's just seeking the path of least resistance to end up as

14:49low as it can

14:49go, because of gravity, right? So with AI, think of AI as just water. It's like, oh, it finds a

14:56path, it's going to go

14:57around it, it's going to pool it. If it finds a pool, it's going to fill the pool until it

15:00spills over the edge

15:01and a little waterfall and all that kind of stuff, right? So you just give AI a goal in the

15:06same way that

15:07water has the, quote, goal of getting to its lowest possible path before it can't get any lower. It wants

15:12to get to the center of the earth and float there and air beads or whatever. And so that's what

15:17water is

15:19doing, that's kind of what AI is doing. You give it a goal, and it's just going to try and

15:22find every

15:23conceivable way to get to that goal, and that's what it was doing. OpenAI said the models did not

15:28find an easy, unmonitered path to the open internet. No uncontrolled outside compromises were reported

15:35from this exercise. It also led to the discovery of an unrelated bug in a widely used compression

15:39library called Brotley, also known as the sound when you eat too much bratwurst. So humans are using AI

15:47to facilitate hacks, right? And so ConnectWise Storm0940, which sounds like a neo-Nazi handle.

15:54So a China-linked group used open source tool that lets AI systems plan and carry out multi-step

15:59actions with limited human input, right? So you basically just, you take the AI, and basically

16:03it's like you're dumping water on the top of the mountain. It's going to find its way down. You say

16:07to AI, just find your way in to this system. In one phase, they tried common passwords against hundreds

16:12of accounts across eight organizations. Microsoft saw no successful logins from that phase. So

16:19a boomer passwords, right? Boomer passwords, you know, your birthday, or ABC123, that kind of stuff,

16:28right? So after compromising a backup product from ConnectWise, they used the AI tool to attempt

16:34logins against more than 22,250 Microsoft 365 customers. Peak rate was about 2,900 attempts

16:42per minute. Most failed, but in some cases, they gained privilege access in under three hours.

16:47Some secondary reports say successful unauthorized logins reached the low hundreds of tenants. The

16:52ConnectWise product itself was compromised. Downstream success was partial, but real.

16:57Latin American financial institutions, aka money laundering. No, I'm just kidding.

17:04A little bit. All right. A group targeted more than 400 banks, payment apps, and other financial

17:08institutions in Latin America. They used commercial AI models from Anthropic, Claude,

17:14XAI, Grock, and Moonshot, Kimmy. Research from the firm Straker indicates passwords and customer

17:20data were stolen, and individual bank accounts were accessed in at least some cases. An internal

17:24database held thousands of credential records. And don't you just wake up every morning? You wake up

17:30every morning in this lovely, glorious internet planet that we all inhabit, and you wake up every

17:35morning, and what do you get? You need to log into this DocuSign. Oh, somebody ordered AirPods on your

17:44behalf. Your girth can be increased. Terms of service have been updated, and we can now listen to your

17:52heartbeat through your gonads. And you need a password reset. You know, all of the general

17:59fishy, hacky nonsense, right? So that's your... And also, the other thing, I guess, is sort of somewhat

18:06relevant to what we're talking about here. The other thing is that you hear, or you get, your data was

18:12compromised at a hack. We're really sorry about that. We don't really know what happened. We're still

18:16trying to track it down, but yeah, your data's out there. Sorry. Sorry about that.

18:25I mean, there's things you can do, right? So you can buy these products that sort of try and figure

18:31out what information is out there, what you need to change, and so on. And you can also buy products

18:36to get you off email lists, and get your data deleted from external systems, and so on, which may

18:41be worth doing. So hacks are still happening the old-fashioned way. Quest Diagnostics Incident

18:48reported 23rd October, 2026. Boy, it was current when we first did the research, man. No, no evidence

18:54links this to AI tools. Patient data was exposed. The company serves a large share of U.S. hospitals

18:59and physicians. Investigation is ongoing. A former employee publicly described decades-old service,

19:06or decades-old service, weak recovery practices, and loss of knowledge after layoffs.

19:10You know, I'm going to go out on a teeny tiny limb here. A limb so small, Japanese and slender,

19:19that I'm sure by saying this, I'm going to get more emails about how to increase girth and length. But

19:26I will say that taking employees with years or decades of coding experience in a particular environment,

19:34and getting rid of them and getting diploma mill questionable abilities from overseas is probably

19:42not helping data security and retention. Now, this is one of these annoying things that I have a fair

19:50amount of experience in. I have a fair amount of experience managing female coders who get pregnant a

19:57lot. Very exciting. Very exciting indeed. And I also have experience hiring people with extremely

20:04questionable overseas credentials. I'm not blaming anyone for that. I was responsible. I've interviewed

20:11like a thousand people. I've hired like a hundred people over the course of my career, mostly, mostly

20:16pretty, pretty positive. But yeah, I remember hiring a guy who had a PhD from a Chinese university.

20:25Now, this is before it was really possible to look these things up and validate or verify them.

20:30I remember one exciting time when we had to cross the border. And that was a lot of fun. Anyway,

20:38and he was, he was terrible. PhD in computer science. Guy could not write code to save his life.

20:46The most successful people I ended up hiring were either students from the local university,

20:51but we had a good relationship or the people I really liked. I designed my own intelligence test

20:57and just had people pass that. And if they were intelligent, they could do it. But the hobbyists

21:01were really great. The people who were just like, oh, I've been taking, breaking apart computers and

21:04building them since I was 12. And I learned how to code when I was nine. And, you know, I've

21:08loved it.

21:08And those people, I mean, you weren't giving them a job. You were paying them for something they do for

21:13free. You're paying them for something they pay you to do. All right. Or to be able to do.

21:21Unauthorized access was confirmed to the quantum, sorry, quantum lab services manager system,

21:26the portal some clients use to order tests and retrieve results on 21st October, 2026.

21:30The company said core business operations were not disrupted. Excellent. Yeah, that, you know,

21:37when someone steals your health data, the most important thing you're most concerned about is not

21:41who has access to entire medication history potential or your entire medical history.

21:46Or is it John McCain? They allowed reporters into like 400 pages of his medical stuff. Yes. Oh,

21:54maybe all of my medical information is out there on the internet. But the most important thing that

21:58I'm concerned about is whether your core business operations are not disrupted. That's what matters most

22:06to me. The full scope of patient data involved is still being determined. So how are how are AIs used

22:14in the discovery of vulnerabilities? Well, they ask you about your mother. So a brief on responsible

22:21disclosure. So let's what do you have to do if there's a data if there's a data leak, a data

22:27leak

22:28that seems to be happening so often that you're like a seven year old with a prostate the size of

22:31a

22:32watermelon. So a researcher privately notifies the company that makes the software of a security hole

22:37with enough detail to reproduce it before telling the public. The company is given a set period after 90

22:42days shorter if the hole is already being used by attackers to fix it. Public details are released

22:47only after a fix is available or after the waiting period ends if the company does not act. Of course,

22:52as you

22:52can imagine, the goal is to limit the time when attackers know about the hole, but defenders do not yet

22:56have a

22:57fix, right? So open source PHP to shell, two critical holes in PHP, the language that runs a large share

23:07of

23:08the web. These allowed an attacker to run code on a server without needing a password. Yeah, once you can

23:14run code on the server, the server is and of course, the server then spreads the code to other servers,

23:21and

23:21it's a giant mess. So these issues had existed since 2012. But of course, I only discovered them

23:28in 2011. I didn't never discovered them. They were found by AI models, one team using OpenAI's model,

23:34another independently confirming with Google's. An emergency patch was issued on October the 21st,

23:392026. Details leaked before the patch was ready. CloudFare saw exploitation attempts jump from

23:45thousands to millions within hours of public details. PHP still powers roughly 70% of websites that use a

23:51known server-side language. This is the Brotley one I mentioned earlier, a critical bug in a widely

23:56used compression library found during OpenAI's Red Team work. The compression library is also what

24:03I need if I'm not wearing boxers. It can allow problems when software unpacks untrusted compressed

24:12data. Patches were released for the affected versions because many projects bundle their own copy of the

24:16library. Each one must be updated separately. Telegram sticker vulnerability. Ah, Telegram, I have

24:24such ambivalence. A researcher used AI models to find memory corruption bugs in the open source library

24:29Telegram used for animated stickers. Wow. He developed a proof of concept that could run code on iOS devices

24:35without the user interacting with the sticker. You know, I don't mean to sound mean, and maybe I'll get hit

24:45by this

24:45one day myself, but I gotta tell you, I do not have an enormous amount of sympathy for people who

24:51are like,

24:51oh, I, I, like the John Podesta stuff, I will, I will click on this link. It looks like Google,

24:57I will click on this

24:58link and reset my password. I mean, please, everybody, I mean, just get your virus scanners, just, you're gonna have

25:11to give up some CPU cycles. However, the idea that something can hit you without you even having to interact

25:17with it, that's creepy. And of course, what can you do about that, right? I will just, you know, one

25:23of my personal

25:24habits, right, rightly or wrongly, I think rightly, I'm a little obsessive about updating software,

25:30hopefully to the right version that's right. So every couple of days, I'll be like, hey, any new

25:35software? Great, update it. I don't wait for this automatic thing. Just, I assume that everything's

25:40getting buggy and it's got holes in it and they're patching stuff all the time. So I'm a big one

25:46for

25:47patch, patch, patch, update, update, update. The moment, I mean, I'll check every couple of days,

25:52even for operating system upgrades. So I just assume that everything is constantly getting Swiss

25:57cheesed through AI, and they're constantly putting out patches. So just my particular advice, update,

26:03update, update. Hopefully not to something that's vulnerable, but I assume that there's some reason

26:08for, you know, because they all, you know what they always say, they always say, you know, we've fixed

26:11this, we fixed that, we've improved this, we've improved that. Now, it's semi-liquid glass that, you know,

26:16caresses your nipples in just the way you like. And then right down, it's like, also massive security

26:20vulnerabilities have been passionate about that kind of stuff, right? So Telegram had claimed

26:25earlier mitigations the researcher worked around them in early September 2025. Six, Telegram replaced

26:30a vulnerable library in its clients. Other long-lived bugs found by area systems include a 15-year-old

26:36local privilege escalation hall and an 18-year-old virtual machine escape. Critical vulnerability

26:42reports have risen sharply from a few dozen per month a few years ago to roughly 600 per month now.

26:48So it's, it's like all AI. It's mostly good with some bad. AI is going to be really good at

26:54finding

26:54holes. AI may be helpful in patching them. These holes have existed for a long time and it's kind

27:00of better that they're found. Now, one of the problems is that if there are hardware problems,

27:04then you have a bit of a, bit of a problem. All right. So closed source, this was all the

27:11open

27:11source stuff. What about closed source? Anthropics, Claude, Claude. I always think of that as a

27:17galatoire smoking, a bittered guy in a striped shirt holding a baguette with a beret. Claude in a

27:24research project. He's a racist. Worked only from compiled Windows binaries, the machine code without

27:30source and produced working proof of concept crashes and full privilege escalation chains. So full

27:36privilege is when you can do anything and everything, right? So I don't know if you've ever had a lockdown

27:41machine. Sometimes this is the way it works. You can't really mess around with the settings too much

27:45and so on. They sometimes will do this when they have computers out. Best Buy or other places like

27:51that to make sure that nerds don't come in and fry their machines. So other systems have found dozens

27:58of previously unknown holes in ordinary Windows components, with some confirmed by Microsoft and

28:02resulting in bounty payments. Yes, Microsoft has, as a lot of places do, bounty for bugs. Dating when a bug

28:08was first introduced is harder without the source code. So I just wanted to sort of mention that.

28:14Can AI help with fixing security vulnerabilities? Can it get people to stop asking me about my mother?

28:20A 2026 report that tested more than 100 models over four years found that on average,

28:25AI-generated code still introduces a detectable security problem in roughly 44% of tasks when

28:30given no special security instructions. The rate has stayed flat, even as models have gotten better

28:35at producing code that runs. Larger or specialized coding models have not shown a clear advantage on

28:40security. Functional correctness has improved, security correctness has not. The volume of AI-generated

28:46code has risen sharply, while the rate at which it introduces vulnerabilities has remained roughly

28:52the same. So how likely is it that old vulnerabilities were exploited? A 2017 RAND study,

29:00the corporation, not the writer, of more than 200 real-world zero-day holes, holes unknown to the

29:07software maker, found an average life of 6.9 years from private discovery to public disclosure.

29:12About a quarter of these holes lasted more than 9.5 years. That's longer even than Sting's tantric

29:19sex sessions. Or maybe. I couldn't say for sure. I'd have to check my feet. Once a hole becomes public,

29:25the time until attackers start using it has fallen sharply, from roughly two months a few years ago

29:29to a few days in recent data. Even after a fix is available, organizations often take weeks or months

29:36to install it. It is not possible to prove that any specific long-lived hole was never found and used

29:40privately before it became public. A careful group could have kept, you know, if you find a

29:44vulnerability, you keep it secret, you use it sparingly, and so on, right? Like that guy,

29:50I guess this was used in office space, who took a quarter of a penny in interest calculations and

29:56rerouted it to another account. You want to keep it small so it can't be seen that much, and don't

29:59use

29:59it too often. For well-studied cases such as the PHP holes, the lack of earlier public science of

30:05widespread use is meaningful evidence against routine prior exploitation, but limited high secrecy use

30:10cannot be ruled out. Uncertainty is larger across the full range of software. Most long-lived code

30:16has not previously received the intensity of review that is now possible with AI-assisted analysis.

30:21And the other thing too, I don't know if you've been following some of this stuff, but in my humble

30:25opinion, you kind of want to check your televisions, because it seems like a fair number of televisions

30:32are listening in and listening for words. They're listening for babies crying or dogs barking or

30:39things like that, so that they can serve you up targeted ads. Maybe they're doing more than that.

30:43I don't know. But most of your systems have the ability to opt out of that. So boot up your

30:50TV,

30:50go to your settings, and opt. I mean, I just recommend using Android, using iOS, using Windows,

30:56whatever it is. Linux is a little better. Just ask AI, what are the security holes that are in the

31:03technology around me? And wherever you can, just try to opt them out as much as possible. So that

31:09would be my suggestion. Some of the TVs even seem to, if they don't have a microphone or if the

31:14microphone has been disabled, they'll look for attached devices to them that have a microphone

31:19and turn that on instead. Just take my spleen, man. Just take my kidney. Just do it. And that's

31:27one of the reasons why your TVs are so cheap, because they've been stealing your soul and selling

31:30it to gladvertisers. All right. So what is going on with AI? How much is being spent on AI these

31:38days?

31:38So Anthropic reported $16.3 billion in revenue for the first half of 2026, up from $1.1 billion in

31:45the

31:45same period of the year before. And a large accounting loss driven mainly by a one-time

31:49write-down. Underlying operating burn was roughly $8 billion. OpenAI has roughly $50 billion in

31:58annualized revenue and reported a large operating loss in 2025. Free cash flow is projected to stay

32:03deeply negative through 2030. Amazing, eh? To make $50 billion in annualized revenue and still be

32:14deeply negative is really something. Computing is the dominant cost. Anthropic spent nearly $9 billion

32:20on compute power in the first half of 2026 alone. Long-term locked-in infrastructure commitments

32:27exceed $600 billion across cloud contracts, chip leases, and related deals. Frontier model prices have

32:34risen while prices for non-frontier models have fallen. Anthropic is preparing a large IPO amid investor

32:40questions about growth and financing costs. So a lot of people, of course, are saying that AI

32:46is a part of the boom-bust cycle that hits tech and mostly tech. And I was around for the

32:52one

32:52in the late 90s and early 2000s. I got out of my tech career and took time off to write

32:59right before the tech crash, which was, I mean, I'm not going to claim that it was any kind of

33:05wisdom.

33:05I just happen to get lucky. All right. So the late 90s, early 2000s technology reset had two linked

33:11parts. Demand-side internet company valuations collapsed after the March 2000 peak. Many firms

33:17with little revenue could not turn usage into sustainable income once financing dried up.

33:21The, of course, infamous example of things like pets.com, where you could get pet food delivered

33:25to your home. So on the demand side, the customers just weren't ready. And the internet was pretty

33:34primitive back then. It was kind of tough to use. On the supply side, telecom companies spent roughly

33:40$444 billion on fiber and equipment from 1996 to 2001. At the peak, this was about 5% to 6

33:48% of the

33:48global information technology market. Much of the fiber sat unused for years. The result was large

33:53bankruptcies, sharp cuts in spending, and hundreds of thousands of job losses. Current 2026 capital

34:00spending guidance from Amazon, Microsoft, Alphabet. And Meta is $720 to $750 billion, mostly for AI

34:07data center capacity. Against the roughly 6.4 trillion worldwide IT spending forecast, that's

34:14about 11 to 12% of the market, roughly twice the relative intensity of the peak telecom year.

34:19The companies building the frontier models are mostly renting capacity rather than building it

34:23themselves. Their multi-year cloud and chip commitments total hundreds of billions and flow

34:27to the hyperscalers. They are not additional physical capital on top of the hyperscaler numbers.

34:32Model company revenue is growing very fast, but it's still far below the scale of the locked-in

34:36infrastructure commitments, right? So if somebody's going to build you a giant server farm, they want

34:41you to lease it for a while, right? They're not going to build it for you to lease it for

34:45six months,

34:46because then they're going to be left with a whole lot of hardware and not a whole lot of video

34:49games to play it on. So the outcome depends on whether high value usage materializes at the

34:55levels assumed in the capital plans. I mean, obviously there's the consumer fun party stuff.

35:01What was it? Somebody did the Fantastic Four as Bruno Mars and Mick Ronson and the others for Uptown

35:12Funk. This one for this poor girl, straight masterpieces. They had all of the Marvel characters. I think it was

35:18Marvel characters doing it. It's kind of fun. But where AI is going to really matter is in terms of

35:27replacing people, right? Whatever allows you to displace people and fire people is something that

35:34is going to make money in the long run. So there's going to be a big battle, of course.

35:41Without a doubt, AI can replace HR for the most part. Again, some, right?

35:47AR can replace a lot of accounting. Sorry, AI can replace a lot of accounting. AI can replace

35:54a lot of law. AI can replace a lot of, you know, anything that touches keyboard, screen, and data,

36:02AI is going to tear through. A friend of mine is an excellent, excellent businessman,

36:09has, well, no longer needs to hire business consultants because he can just have AI look at

36:14his business and make suggestions. I mean, so there is going to be a lot of replacement.

36:19The first wave, of course, is going to be digital work. And the second wave is going to be manual

36:27labor or skilled labor for that matter, right? So yeah, it's pretty, pretty cool.

36:33So the resemblance

36:37is large physical infrastructure spending justified by expected future demand,

36:41occurring alongside heavy promotion of the technology. In the late 90s, weaker than expected

36:46demand and overbuilding reinforced each other and produced a significant contraction in technology

36:51investment and employment. Of course, it comes down to whether actual usage tracks the levels

36:57assumed in the plans. It is a risk to watch, not any kind of prediction. So what do you guys

37:03use

37:04for AI? Just out of curiosity. I'm a bit of a Grok guy myself. I've tried a couple of others.

37:09Not bad,

37:10not bad. But I think I've tried Gab AI and Grok is good because I don't think that the Grok

37:20engineers

37:20are intercepting everything to turn it super deadly claustrophobic woke.

37:26I mean, one of the big tests I have of AI is just ask it questions that would make people

37:34uncomfortable, right? That's all I'm looking for. Just have it ask questions that people would find

37:39uncomfortable. Just ask about AI and things like that. And just, sorry, ask about IQ and things like

37:44that and just see, see how it does. And a Grok is, uh, it's pretty good. It's pretty good that

37:50way.

Category

Transcript

Recommended