340 | Cyber-geddonm! OpenAI Disconnecting Training from the Internet 🤯, Microsoft is Buying Out 60K employees, AI is taking over the commerce world, and more important news for the week ending October 30, 2026

0:00Speaker Hello, and welcome to a weekend news episode of the Leveraging AI podcast. The podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Meitis, your host, and we had a week that seemed to be quiet for a little bit, and then a lot of crazy stuff happened.

0:20Speaker Some of it is directly related to AI, some of it is not directly related to AI, but is directly connected to the bigger picture. And what I'm talking specifically in this particular case is the breach of Binance, the largest crypto exchange in the world. Nobody still knows if AI was involved or not, but it connects to a much, much bigger story that we were touching upon multiple times in the past few weeks, is the rising risks of cybersecurity, and then a few other things came on the same exact topic.

0:48Speaker So we're going to open today by talking about cybersecurity again. I'm gonna do this in an interesting way. I'm gonna tell you The timeline from the Hugging Face incident that maybe made this very real for everybody who's paying attention all the way to what happened in the past couple of weeks, and definitely the last few days.

1:05Speaker And what you will see is that, A) it's becoming from, "Oh, these agents are randomly, without intentionally, actually breaking into all these places," to, "Holy crap, hackers are actually using this to do some very serious damage," and what's happening behind the scenes at the same time. So we have a lot to cover on cybersecurity, including some news from OpenAI that found that even when they're trying to prevent AI agents in testing to getting to the internet, they can still do it and figure out ways to do this when they use virtual machines, and what they're going to do about that.

1:36Speaker So a lot to cover basically about cybersecurity. But then we have two additional interesting topics to dive into. One of them is the ROI of AI and where it stands right now on the bigger macro picture, not at an individual company. There's a lot of interesting news, and again, I'm connecting the dots from multiple sources for you to tell you a broader story of where we stand right now.

1:58Speaker And then we're going to talk about AI in commerce, because a lot of things are happening, in the last few weeks that are all hinting that we are rushing into an era where AI is gonna do more and more shopping for us and how that looks like from the current status, how does that look like from an infrastructure perspective for some of the leading companies in the world?

2:20Speaker What is happening in this field at all? Again, there's a lot of interesting news related to this that I'm gonna combine for you in order to paint a better picture of all of that. And then we have a lot of interesting things to talk about in the rapid fire. first of all, Anthropic has formally filed for an IPO, so they actually submitted, uh, registered their statement for, their initial public offering.

2:41Speaker And then we have a new extremely powerful model from Google that is taking over the charts, and we have new charts, which is something again, I'll explain later on. So basically a lot to cover. So let's get started. Now, before we get started This episode is brought to you by the Multi-Agent Orchestration Course.

2:59Speaker It is a course I have been teaching personally in a live cohort since May of this year. Every single one of them sells out, and this one is the first one that didn't, so there are a few seats available. It is starting this coming Monday, meaning October 2nd. So if you're listening to this episode still this weekend, you can still join us.

3:18Speaker as I mentioned, we just have a few seats left. In this course, we'll teach you basically how I'm automating my entire two and a half businesses. The exact infrastructure that I'm using, the exact processes that I'm using, the exact methodology that I'm using, step by step on how I'm building everything that I'm building for myself and for my client, I'm going to teach in that course.

3:37Speaker And again, it's actually me. And if you want to join us, there's a link in the show notes, and you can still go ahead and do that, and you can use leveraging AI 100 to get $100 off the course. With that, let's dive into our first topic, which is cybersecurity. And I'm gonna take you back. Again, this is gonna be a lot of timeline.

3:56Speaker So we're gonna talk a lot of dates in order to prove my point that things are accelerating like everything else in AI very, very quickly. So on May 7th, OpenAI starts reinforcement learning of a new experimental model On May 8th, they're giving it some task that is practically impossible just to see what's gonna happen, and the agent is trying to attack OpenAI's internal software repository.

4:19Speaker It fails, but the agent learns from it, and it understands some of the things that it can do as far as doing things that it didn't know it had access to before. From May through July, Agents that are being tested by OpenAI are using a dormant German wiki as a messaging board, and they're using about 18,000 edits on that wiki board in order to communicate between the agents back and forth So again, just to explain, this is not something the agents are supposed to have access to.

4:49Speaker They just found a place where they could communicate without being detected. Again, they made 18,000 edits in those two months, while in the previous 10 years, there have been 20 edits to that wiki. So they found a website that nobody's watching, and they started communicating through that website. On June 18th, the agents get into Australian Medicare as well.

5:09Speaker On July 9th, agents are starting to leave the test environment, and they start doing more and more things that allow them to do more and more stuff outside of their sandbox. July 11th through 13th is when these agents actually hack Hugging Face, and they do this by using swarms of agents that do multiple things Hugging Face themselves counted about 17,600 actions on their website with over 1,200 agents in order to get access to what they were trying to get access to And on August 18th, OpenAI pauses the reinforcement learning training, uh, for two weeks, in order to go and investigate what's happening.

5:49Speaker As I mentioned, on September 24th, Australian Prime Minister announces that their Medicare website has been breached again by AI agents, by OpenAI. So this was just what made everybody aware that this is actually happening. Since then, we're hearing more and more of similar events that are happening across the board.

6:08Speaker So on September 20th, during a training run, an internal research model used a misconfigured DNS server at OpenAI to get access to the live internet Again, this is not a model we know. OpenAI didn't tell us what model it was, but they paused the training again in order to verify what's going on, and it took them 11 days before they started this again.

6:31Speaker This is after everything they've learned or presumably learned from the Hugging Face incident. The same exact thing happened to them again on September 20th What OpenAI did in order to figure out how to be safer in the future is they've spent hundreds of billions of tokens based on the Nextweb publication to create a AI-based red teaming campaign to try to get agent swarms, uh, running against isolated copies of its research infrastructure.

7:00Speaker So they basically was trying to see what happens if agents attack OpenAI's own infrastructure by creating replicates of that infrastructure and letting the agents try to attack it. What they found, and they shared that in a publication on October, 23rd, is that agent chained software bugs to escape virtual machines and run code on the host machine Now, I'm not going to cover the exact technical details because they mean nothing to me and probably mean nothing to most of you.

7:26Speaker But the bottom line is these agents were able to find really sophisticated, really complex steps in which they can jump through loopholes in different systems that is supposed to keep them in place Now again, things that are not supposed to happen, which we all do understand. They were testing separate agents in separate sandboxes, but on the same hardware.

7:47Speaker The agents found ways to communicate from one sandbox to the other, passing secrets to access different things from one sandbox to another sandbox that again, not supposed to be possible. And they did this through processor timing differences. There was no network, no shared files, no nothing, and yet the agents, by using the actual hardware itself, were able to communicate with one another Now, what they are saying is that they've since then patched a lot of the issues that they found in the software of the containers and in the setup of their containers, and they're saying that it has significantly dropped the number of findings in this environment where AI can actually breach.

8:29Speaker However, they're also saying that while they now have a much more secure environment than they had before, they do not know if they'll be able to catch more advanced models doing things that they don't even understand Now, on October 22nd, OpenAI also shared that Meter and Redwood Research both had external, so these are third-party companies, they had access to their environment in order to test and monitor their agents, and they found that Codex agents, and I'm quoting, "Could in principle bypass Codex sandboxing and auto review."

9:02Speaker And that about 10% of internal Codex traffic went unmonitored from May to July because of a gap in the systems and the way they were set up. So again, 10% of the AI that was working on Codex wasn't monitored to see what it's doing and if it's doing anything malicious or just wrong without being malicious.

9:22Speaker So on October 23rd, a day later, OpenAI changed their policy, and they're saying that their frontier model training now runs with no live internet, period. It's not a software container. It's actually a separate machine that cannot technically be connected online they're doing this obviously, and I'm quoting, "We must always be able to prevent a frontier model from achieving unauthorized arbitrary access to the live internet."

9:50Speaker So they're going from we build these really powerful containers in which we can test the models to we know we can control the models, so we're going to test them in a box that is physically not connected to the internet Now, a similar external test was done by Anthropic working together with the UK AI Security Institute.

10:10Speaker In that experiment, nine of 43 runs, Claude Mythos 5 went beyond the exercise Now what this model did, it was trying to learn how to do more things using GitHub, and in order to get access to what it needed in GitHub, it was trying to convince an owner of a repository to plant malicious code into the repository so it can take over it, and it was trying to convince that person that it is a different person and get access that way But all of this, we're talking about agents that are doing stuff on their own without actually being directed by people to do malicious stuff.

10:46Speaker These are just agents going rogue, trying to solve problems and achieve goals that they were set. By the way, the concept of OpenAI wanting to run their training runs without internet connection means the researchers needs to physically be at the data center, because otherwise they can only access, these agents through or these tests through an internet connection, which is what they're trying to prevent.

11:12Speaker I'm not exactly sure how that's technically possible, but that's what it means. At least that's what I think it means. It will be very interesting to see how OpenAI actually pulls this off So now let's talk about more recent cybersecurity issues that are all actually done by bad actors who are using AI, in some cases we know that, in some cases we're not 100% sure, to do really bad things.

11:35Speaker On September 25th or between September 25th through mid-October, as far as we know, a Chinese hacker that used a very short set of instructions, less than 100 Chinese characters, attempted more than 1,200 sites breaching, of more than 100 companies, including some big commerce companies. Now, nobody knows exactly what information was stolen, if information was stolen, but what we know is that he used about $5,000 worth of tokens for the entire process, which is about $50 per breached website.

12:09Speaker That is practically free if you're gaining some important information, and this is just showing how capable these models are becoming right now and how scary the situation is, where hacking capabilities that previously existed to a very short list of people around the world who would focus their efforts on very specific things, now more or less anybody with relatively basic knowledge can apply really advanced hacking skills in order to get access to a very broad range of websites In early October, Shinhan, KB Kookmin, and Hana, which are all South Korean banks were attacked by a Chinese hacker that use a platform called Artex, A-R-T-E-X, which is an agentic, tool, plus Claude Code sessions in

12:57Speaker order to breach those banks. This information comes from CrowdStrike Customer data and credit card information were exposed, and some systems were down. Again, this is early October. Still in October, a single actor ran 43 AI agents through a framework called CyberStrike AI, and at least 19 confirmed victims and ransom deployed, all coming from a single hacker, a single source that was applying all of those.

13:24Speaker On October 17th and 19th, uh, attackers stole API keys from 1.4% of users of Coding Tools Console Some users saw charges of above $1,000, and OpenCode's defensive AI models refused to keep, trying to analyze this, very similar to what happened with Hugging Face, and so they had to go to open source models in order to block, uh, the attack.

13:47Speaker But in this particular case, if you think about it, this is a really sophisticated kind of approach where you're stealing the API keys of somebody else as your first step, and then you're getting access to free or cheap tokens, and then you can apply bigger attacks by using other people's accounts. So it's very hard to detect who you are, and you're doing this for a lot less money because you just stole a big chunk of people's API keys, and you can use their AI instead of your own.

14:12Speaker And then October 21st through 22nd, over 400 banks and financial institutions in Latin America were attacked with a swarm of agents This was again a mix of Claude instances that planned and decided what the attack is going to be, Groq agents that scouted the targets, and Kimi, plus some conventional malware that existed in the market was used when Claude refused to do whatever things it was asked to do.

14:38Speaker So operators tied together, or these hackers tied together multiple different agents, played each and every one of them to its maximum while overlapping between them in order to avoid the places where the models either weren't good enough or would not participate in the bad thing that they were trying to do We don't know everything that was stolen, but a database that held over 4,400 records with emails, account numbers, passwords, API keys, credentials to the banks and so on was stolen.

15:07Speaker Uh, we also know that additional financial sensitive data was stolen from at least one unnamed institution Now, we know all of this because the group of hackers that did this actually leaked the operational data themselves, and then somebody else found it and shared it on the internet. So all of everything that I'm telling you is coming from the horse's mouth in this particular incident So what this is showing us is that while in the initial cycle of this, we had AI agents going rogue and trying to hack different things just because they're trying to achieve their own goals that were set to them in training, what we have right now is real bad actors doing real bad things on their own because they

15:42Speaker can in ways that they just couldn't do it previously. Which leads us to the whole PHP to Shell issue that we reported about last week. So on October 17th, Amazon Honeypot network logs an exploit attempt, just four days before the public disclosure of the issue with PHP, which means people are starting to figure this out.

16:04Speaker If you didn't listen to last week's episode, uh, this issue was found by two different company, Calif and Wiz, both of them are cybersecurity companies. They found this and worked with the government, the NSA, and other government agencies, and I assume some large other non-government organizations, just patching, uh, these issues.

16:22Speaker But the status leaked out, which drove attackers to immediately go after these things. On October 21st, PHP themselves shipped an emergency patch, uh, a day early, uh, of the announcement for several of its versions And then the NSA came out with its announcement saying everybody has to patch these things otherwise everybody's at risk.

16:43Speaker Now, another interesting thing, Cloudflare, which monitors the traffic across a big chunk of the internet, showed that thousands of exploit attempts in the 10 hours before the announcement by the NSA happened. So again, thousands before the announcements. Tens of thousands within 45 minutes after the announcements, hundreds of thousands within two hours, and millions by the time that Cloudflare did their post just a day and a half later.

17:09Speaker So what Cloudflare did is they turned protection to its customers, But since PHP sits on the infrastructure of over 70% of websites on the internet, the vast majority of the internet was still at risk because most people were not patching things fast enough. Now, the speed of scale tells you how scary AI cybersecurity can happen.

17:29Speaker From thousands before it's announced to millions within a day and a half later. This is not something humans can do, but it's definitely something AI agents can do because you can spin as many of them as you want and attack as much things as you can before it gets patched That wasn't the last big thing that was found.

17:46Speaker It wasn't as big, but definitely another big warning. OpenAI red teaming using a model that they did not disclose found a critical flaw in Brotli, which is a data compression library that is being used in many different places, such as Node.js and .NET and other software tools that use Brotli data as a way to, pack and unpack data.

18:09Speaker Microsoft shipped an emergency .NET update on October 25th And on October 27, CISA added it to a list of actively exploits that needs to be addressed immediately. So what we're seeing is we are now in a race that happens most likely every single day between the good people who are trying to use AI to find vulnerability and exploit ways and block them and patch them, to the bad people who are looking for ways to use them, those exploits, before they're getting patched.

18:42Speaker My gut feeling tells me is it's not a race that the good people can win 100% of the time. Meaning, yes, we can probably block a lot of things, and a lot were already blocked, and yes, the labs themselves have more advanced models than they're releasing to the public, and they're giving it to the right people to try to find these vulnerabilities and patch them.

19:01Speaker Just to put things in perspective, some of these vulnerabilities have been around in that code that is infrastructure to, again, the entire internet or in cases just servers and stuff like that, for 10, 15, 20 years without being detected by humans, and now AI can find them very easily, more scary, find ways to exploit it and do really bad things with the fact that it can be exploited until it gets patched Now, another event that happened this week that may or may not be AI related, but it's just showing you how problematic the cybersecurity situation we're in right now is that Binance, which is the world's largest crypto exchange, got hacked.

19:41Speaker Binance estimates that $6.3 billion were impacted. Uh, they can't say exactly when or how or where it came from, But what they know is that some things were hacked by using a DevOps engineer compromised laptop. So somebody was able to hack not the actual users themselves, so not the people who hold the cryptocurrency, but a person with high entry capabilities into the system because he's an engineer, and through that get access to what is supposed to be unbreakable.

20:17Speaker So the whole idea behind crypto is that you can't hack it. Well, apparently, if you got the right privileges of a developer, then you can do a lot more things than you think, and that just happened this week What did Binance do? Well, first of all, they froze about $3 billion worth of crypto right now, and they've suspended withdrawals for a while.

20:40Speaker They're saying they will cover the losses through a reserve fund that they have. It is not exactly clear if they have enough funds in the reserve, uh, right now to cover everything that might have been... And they're not saying it was all stolen. They're saying it was affected. I don't exactly know what this means other than stolen, but it may just be a way to say this in a more politically, correct and less harsh way.

20:59Speaker That sent Bitcoin down 4.8%. Which is, again, an obvious outcome when you think that your unaccessed funds that is on the most secure network ever created to secure funds was now hacked and people can get access to it Now, before I give you a quick summary of where I think this is going, the one last thing that was found is Anthropic Frontier Red Team says that an unreleased internal model with only initial prompting, so nobody's trying again to do anything severe, Found vulnerabilities in Chrome, Safari, Linux, Windows, Android, and so on.

21:36Speaker And what they did is for each and every one of these issues, they created a detailed report on how to patch it and send it to the relevant vendors so they can patch it before, uh, it gets released to the public and before people find out about it. But again, think of what we're saying. This is what we managed to find.

21:53Speaker These are the core infrastructure that everything we know runs on. Linux is most servers in the world. Safari and Chrome are the browsers that everybody are using. Android operating system is on most of the phones on the planet, and so on and so forth. So every major infrastructure, literally every single one, has vulnerabilities in it that were very hard to find and even harder to exploit until now.

22:19Speaker So my real personal fear, and I've shared that in previous episodes as well, that something really bad is going to happen sometime in the next X number of months. I don't know if that's three, six, 12. And what I mean by really bad is once bad actors saying, "Ooh, we wanna take action that will be dramatic, and we wanna do this before it gets blocked and patched by whatever AI tool," they may pull the trigger on things that could be catastrophic.

22:47Speaker And what I mean by catastrophic is it could be really catastrophic that will actually kill people, or it could be just catastrophic from a financial perspective. So think about if somebody shuts down Nasdaq Or just messes with major infrastructure in a major metropolitan in the US. Either of these things will send the stock market crashing, which will have a very dramatic effect on anybody who has their money in the stock market, which is a lot of people.

23:12Speaker So we are in a situation right now where, like I said, we're in this daily battle between the good guys and the bad guys. And while so far the good guys have a step ahead because these unreleased models are finding things and being patched, I don't think we can patch everything. I think that's very naive to think that we can patch everything because before bad people use this to do really bad things, and I'm very uncomfortable with the current situation, but this is where we're at.

23:40Speaker What does it mean to you? it means to you as an individual, update everything. Every time you see an update for your phone, an update for your operating system of your computer, an update for the version of your security cameras, like everything that you can update, go and make sure you're on the latest update and keep updating all the time in the foreseeable future.

24:00Speaker Don't say, "Yeah, whatever. I'll do this sometime," because you might expose yourself to much more serious risk than we ever had before. The other thing that you need to do, if you're a business owner, if someone in business, you need to do a serious estimate of the potential risks that you're exposed to right now.

24:15Speaker You can start by asking Claude, ChatGPT, Grok, whatever tool you're using, Gemini, about what might be the risks that you are exposed to right now as a business, and what can you do to reduce that risk. Check that you have coverage from a insurance perspective for these kind of things. There are insurances that cover you against this kind of risk.

24:33Speaker Is it going to be more expensive? A hundred percent. Is it gonna be worthwhile? it depends what happens. I cannot tell you that. but I'm telling you that it's worth investigating whether it is something you wanna invest as a business, depending on the level of risk that your specific company, your specific industry, is exposed to, and then decide what you wanna do forward.

24:52Speaker But just be aware that this is happening, and that you need to take action in order to reduce the risk as much as possible. The next topic that we're going to talk about is AI or agentic commerce. The thing that triggered me to do this research and give you a coverage about what's going on was initiated by an article or a research that was released by Semrush on October 21st.

25:15Speaker Those of you who don't know Semrush, they are a one of the two most known platforms to do SEO work. And so they released a study. In that study, it's a 12-month study, they looked at Google Overviews, so the AI-based overviews instead of the keywords that people actually click in or the results, the blue links that people click in.

25:36Speaker And what they checked is for 400 tracked US commercial keywords, the coverage of Google AI Overview, again, that doesn't require you to click on any link, rose from 45% or 45.75 if you want to be specific, in September of 2025, it rose to 88% in August of 2026. almost nine out of every 10 results for these 400 specific commercial keywords is now being served through Google Overview, which means you're gonna get significantly less clicks.

26:12Speaker Now, when they took this number, which is 88%, and weighted it by search volume, it got to 97.08%. So basically, almost all the search that went to these 400 words went to get answers on Google AI Overviews versus people being served links so they can click on to go and actually go to your website Now review queries, I wanna know the reviews about X or Y rose from thirty point fourteen percent to eighty-nine percent.

26:43Speaker Again, a huge, huge, huge jump from what happened less than a year ago Now, what they're also showing that in addition to the fact that many people just don't click anything, they just read the AI overview, what Semrush research found is that sometimes the AI will actually send you to the, quote-unquote, "wrong website."

27:03Speaker So the example that they showed in the research is showing, the keyword of Mailchimp versus Klaviyo, and what the Google AI overview did, it cited Klaviyo's own webpage. So if you would've clicked on something, which most people don't, but if you would have, you would've went to the right website. But the information about Mailchimp wasn't actually sourced from Mailchimp.

27:25Speaker It was sourced from a third-party Zapier guide. So even if you would've decided to click, instead of going to Mailchimp, which is the company you wanted to know about, you would've went to a third party participant, meaning Mailchimp would have gotten zero clicks out of that comparison that they were asked about So the way this works, these summaries pull from everywhere they can, YouTube, review sites, comparison guides, et cetera, et cetera, and not just from the source websites themselves, which, A, means the information might be less accurate.

27:58Speaker B, it means that the reference link is not gonna go to the original company's website, which is going to reduce the clicks that you're going to get even further Now, Google themselves confirmed that this is the situation. Google, in their Q3 earning call that was just happened on October 29th, uh, Senior Vice President and Chief Business Officer, Philipp Schindler confirmed that AI mode queries carry lower monetization than traditional search Now this is already slowing the growth rate of search.

28:30Speaker So search revenue is still growing, but it is growing at a slower pace than it did before, which took a swing at Google's actual stock price because it is slowing down, and it is very obvious that AI is going to keep on growing moving forward Now, how is Google going to approach this? They're expanding their ads into AI overviews and AI mode follow-up conversational exchanges, in the US, meaning in the AI summaries, there are going to be ads embedded into that, and in the conversation you're gonna have with Gemini after you get the initial response, there are going to be ads involved into that.

29:05Speaker This means Google themselves are starting to switch their business model that has been one of the most successful business models in history, is changing because of the AI summaries and the way consumer behavior is impacted by those AI summaries That means, by the way, that if you were dependent on organic traffic, that is going to get an even bigger hit, and you will have to learn how to play a completely new game of paid AI traffic, which is something nobody knows how to play yet because it's evolving as we speak Now, the other side of that is that the ad infrastructure itself is turning into an agentic environment, and there are ads that are being created for this new universe.

29:46Speaker So a few examples for that, CNN International had its first agent sold revenue ads. So basically an agent that is selling ads on the platform in Q1 of 2027. Their VP of digital business operation, Faisal Karmali, confirmed that this is their focus moving forward and that they're gonna keep on pushing in that direction But the direction is clear.

30:09Speaker The idea is to allow agents to run your ad spend in the most effective way. There are actually two competing standards in the industry right now. One is called Ad Context Protocol, AdCP, and the other one is called Agentic Advertising Management Protocol, AAMP, and they do roughly the same thing. There is no clear agreed one of them moving forward.

30:32Speaker I assume eventually one of them will win, but not necessarily Another example is Time Magazine They have, as an example, addressed it on the other side of things. They converted all of their web pages into markdown files, so AI can read them easier And they are allowing agents to help in advertising.

30:50Speaker Ally Bank, and Project Management Institute are both advertisers that started using this new capability Another company that is going the direction of letting AI run your ad spend is Reddit. So they now have what they call the AI Ad Guidance tools, and it is a campaign planner that can predict reach, impressions, cost, and so on, all pre-launch, and it can optimize in real time over time as the campaign is running.

31:16Speaker What they're claiming is that advertisers are achieving 80% plus rating and that they see 19.6% lower blended CPA, which stands for cost per acquisition So what's happening in addition to the changes in search is that the way advertisers are going to buy and spend their money is gonna be run by agents that will, A, try to come up with the best campaign, and B, will optimize the campaign as it is running in real time versus having humans do it.

31:47Speaker If you think about an entire industry of ad agencies is at risk with this just one sentence that I just said Now, it doesn't stop at the search level and/or the managing your budget level. There's a whole other infrastructure layer that is involved with this. So Stripe announced a new platform that facilitates AI agents purchases through what they call Link Digital Wallet.

32:12Speaker it is using a single-use virtual cards with spending limits tied to the user-approved purchases and That means that your actual credit card does not touch the actual agent But it also means that Visa and Mastercard or American Express do not get a cut of all these transactions. So if this is the new future, if agents to agents commerce is what's coming, and it is coming, it just doesn't-- we don't know exactly what the scale is going to be, it means that the actual credit card companies may take a very big hit unless they find a way to play a big role, as big as they're playing right now, uh, in this new world of agent commerce versus human commerce To tell you how fast this is growing, uh, Mark

33:00Speaker Zuckerberg, the CEO of Meta and now the owner of Muse, which is, as we reported, the largest number one fastest growing free downloadable application ever. What he reported is a quarter of Muse's 10 million active users are actively using it for shopping and commerce activities with the average transaction size exceeding $200.

33:24Speaker And this is very early in the game Now, I wanna clarify something that I said earlier with regards to the actual credit card. It's not that they don't get any cut, right? Because behind the scenes, uh, Stripe will eventually charge through one of them, but they lose complete control on what's going to happen because they're not doing the actual transaction.

33:43Speaker They're just being, processing some of the money depending on what the agent or the simple card or Stripe will actually decide in the end Also, the people who are going to collect the knowledge about the user's behavior and so on is gonna be Stripe versus the actual credit card companies So what are the two giants Visa and Mastercard doing?

34:03Speaker Obviously, they're not staying silent. Both networks are now promoting their own agent tokenized payment systems. Visa came out with Visa Intelligence Commerce and Mastercard with Agent Pay And they are obviously pushing this as a methodology in a way that will allow agents to pay in agent-to-agent or agent-to-business transactions Now, how much money is at stake in that game?

34:26Speaker US agentic e-commerce spending alone could reach 190 to 385 billion by 2030. That's up from 3.6 billion in 2024, right? So it was practically zero, and it's going to be hundreds of billions within three years. So there is a lot of money to compete for between all these giants and what is going to be the actual payment solution for future agentic commerce Now, while this is happening from an infrastructure perspective, and it is also changes again on the Google side perspective, from a vendor merchant perspective, it is very, very far behind.

35:04Speaker Based on recent surveys done by Visa themselves, only 19% of surveyed merchants have plans, not to mention solutions in place, to accept agentic AI payments If you look specifically at enterprise merchants, the situation is a little better with 28% having plans to do so, again, compared to 19 in the overall, uh, merchant space But this shows you a huge gap 46% of shoppers used an AI tool as part of their last purchase one way or another.

35:35Speaker This is, comes from Global Digital Shopping Index that was done. So about half of shoppers use AI somehow in their shopping process right now, and between 19 and 28% of merchants have plans to allow agents to shop on their website. Again, the 46% is not all agentic shopping. It's me asking ChatGPT to help me come up with what I need to buy or to do research or to decide, uh, what I wanna buy for the holidays and so on.

36:00Speaker But it shows you that the directions on the shopping side is significantly more aggressive than it is on the merchant readiness side. That is a huge opportunity for the merchants that will move ahead faster Now this obviously, in addition to the business opportunity, opens the door for a fraud and identity problem, because if the merchants do not have a means to identify who is actually doing the shopping because it's an agent knocking at their digital door, then that opens the door for a lot of other bad stuff that could happen beyond just losing the transaction Now, there are two things that people can do with AI right now when it comes to using AI for shopping and different kinds of commerce.

36:39Speaker One is the discovery to purchase, right? How do you find the things that you wanna purchase? A company called Feedvisor has done a survey across, potential holiday shoppers. What they found is that 16% of holiday shoppers start their product research with, uh, some kind of AI. Only 9% go to ChatGPT to actually complete a purchase.

37:01Speaker So it's less than 10% versus 16%. Only 5% purchase on an AI recommendation without verifying somewhere else first. So we're still very, very early in that game, but the direction is clear The other part of the process is letting AI do the actual purchase, for you. So a complete agentic process and the willingness to let AI agents act autonomously, look the following when it comes to people in a different survey that I found.

37:31Speaker Uh, book restaurants at 12.1%, buy groceries at 11%, control smart home at 10%, purchase apparel at 7.3%, And so what you can see is that we're still very, very early in the game when it comes to automated shopping. That doesn't take into account, obviously, B2B. I think the situation with B2B, on some cases, it is way more clean and easy to do.

37:55Speaker On the other hand, more risky because of the volume. But you, in the B2B space, in many cases, you buy the same things from the same vendors, and over there it should be a lot easier to actually do. Uh, but it will be very interesting to see how this evolves and how quickly that changes for, let's say, the next holiday season at the end of twenty twenty-seven.

38:13Speaker But as of right now, based on several surveys that I checked, AI is, for most people, the starting point. You do research if you're using AI at all. You do research with AI, and then you still go into traditional review sites and complete the transaction in the traditional ways for the vast majority of shoppers in the next couple of months But what does all of this means to you?

38:32Speaker Well, it means that the world of commerce is changing. Whether you're doing B2B or B2C, discoverability is changing, so the way you are going to be found or your website is gonna be approached is going to change, and it's going to change dramatically. If you are heavily dependent on organic traffic from Google for your revenue, you need to start moving very fast in order to make changes to your discoverability.

38:57Speaker You need to find other methods that will complement the amount of traffic you are getting from Google, because that is going to shrink dramatically if it haven't started doing so already The second thing is you need to start investing in discoverability of your services and products to AI agentic tools.

39:13Speaker There are more and more knowledge about how to do that. That's not the space to do this in this episode, but you have to start implementing mechanisms in place that will allow AI agents to find your services and recommend your services, because more and more of the traffic that's gonna come to your website is going to be agent-based, and without it, you're just gonna lose all that traffic And if you are a merchant, then there are a lot of things that are happening in the background you need to start paying attention to, such as how ready are you to accept payments through these new payments mechanisms?

39:44Speaker Uh, who are you going to work with between Stripe and Visa and Mastercard or all of them? How is this all going to happen? Who's gonna be the main merchant in order to make this work? What kind of ad standards are going to be supported by your website, either ADCP or AAMP, uh, and so on and so forth?

40:01Speaker Will you allow, as a publisher, to allow agents to decide on your spending budget, or are you gonna keep humans in place or some mix of both? All of these things are changing. They're changing fast, and we must, as businesses, pay attention to all of them because otherwise we may stay at the station while the train has already left And now to our rapid fire items.

40:22Speaker The first one, which is the biggest thing that happened this week from an importance to the AI space, is that Anthropic has officially filed their S-1 registration statement with the US Securities and Exchange, for going public Again, this is not huge news because we knew it is going to be coming. What is interesting is obviously the information in there that some of it got leaked when they did the meeting with the big investors a couple of weeks ago.

40:47Speaker But the company reported revenue of sixteen point twenty-nine, billion for the first half of twenty twenty-six, including eleven point six billion for Q2 alone. So that means they doubled their revenue between Q1 and Q2 And it is a 13,000 increase from the 1.16 billion the entire year of 2020, '25 They're also claiming they achieved $974 million in operating profit in Q2, up from $1.7 billion operating loss in Q1 Now, despite these positive things, the company has posted $80.7 billion net loss in the first half of 2026, and these are the numbers they're going public with The interesting thing that I found in the S1 is that Anthropic compute spending fell to 48% of revenue by Q2 of 2026, down from

41:37Speaker 72% in Q1 and down 251% in Q1 in 2025. So What this means is that the percentage of revenue that goes to pay for compute is shrinking, and it is shrinking fast, which means they can become more and more profitable over time as the company grows Now, when it comes to how these stocks are going to be divided, it is not something I'm going to cover today.

42:02Speaker It is really complex. Uh, but Anthropic is going to operate as a public benefit corporation, so similar to how OpenAI is going to be with their, uh, public portion of the company, so their for-profit portion, and they're going to have five different classes of stock at the time of the IPO Now, the most important aspect of all of this, from an ownership perspective is that the founder group, the seven co-founders, are going to own 50.1% of total voting power on key corporate matters, which is making Anthropic what's called a controlled company under the Nasdaq rules, which basically means it doesn't matter how many the shareholders vote against the founders can always make the final call Anthropic

42:43Speaker also committed 5% of its shares to a new public benefit fund They also shared a bunch of risks when it comes to what can happen or what can go wrong, uh, with the stock or the valuation or the company as a whole. One of them is a massive compute commitment. They've disclosed that they have a six hundred and fifty-seven billion dollars in total cloud computing agreement for the next seven to ten years.

43:09Speaker Out of those, four hundred and seventeen billion are non-cancellable. They also have a relatively concentrated customer base risk, with the top five customers representing roughly twenty-five percent of the revenue, uh, for the nine months through September of twenty-twenty-six, uh, with no single customer, though, exceeding ten percent.

43:28Speaker So there is a concentration, but it is not as bad as we heard as far as twenty-twenty-five, where Microsoft and Meta were two of the largest customers and had a very significant portion of the revenue. So it's getting less concentrated, but there's still concentration Now, with all the crazy risks that we are talking about in cybersecurity and so on, obviously Dario was focusing on the positive stuff.

43:52Speaker But if to be fair, he's one of the people that raises both points. But he said, "We believe the diffusion of AI through the global economy represents one of the largest opportunities in history to improve how humans live and work, and it is just the beginning." So amen to that. I really, really hope we're gonna see mostly that and less of the bad things, uh, that AI can bring to the table.

44:15Speaker What we're seeing right now is a lot more on the bad side, or at least the opportunity for the bad stuff to happen. But I really hope that Dario is right and I am wrong Staying on Anthropic, something interesting they shared, uh, Anthropic themselves from a research, is that newer AI models like Opus latest models, uh, so Opus 5.5 and Fable 5.1 can handle complex workflows in a way that allows you to run one big agent instead of a lot of smaller agents, like single-use agents, which is exactly the opposite of what was the recommendation so far.

44:47Speaker So everything that I'm building and actually teaching to build is based on build small agents and small tools that can do specific things and chain them together in order to achieve more complex things. And Anthropic is saying that now a single agent architecture can run approximately twice as fast as a multi-agent setup to do the same exact work with measurably better quality and reduced costs, uh, when you're looking at enterprise deployments.

45:13Speaker Now, what does that mean for every single thing we're doing? I'm not 100% sure. I don't know where are the trade-offs yet, but it is something I'm definitely going to experiment because it is very interesting, because like I said, it is exactly the opposite of everything they and everybody else who builds agents, including myself, has noticed so far.

45:30Speaker So more on that. I may do some experimentation myself and share that as a separate Tuesday episode Now, the place where they said that sub-agents are still a good idea is when the context window is a problem. So sub-agents that can delegate some of the work and reduce the amount of context on the main agent's context window is, uh, still the right place to do this kind of like single main agent with sub-agents that can be doing some of the work and dealing with a big chunk, each and every one of them with a separate chunk of the context that the main agent actually needs Now, I told you in the beginning that we have a whole new way to measure AI.

46:08Speaker so far the golden standard, if you want, the place we all go, myself included, to check how new AI models are doing, how they're compared with other models, is Artificial Analysis. And Artificial Analysis launched a new leaderboard using what they call HLE Diamond which takes into account more advanced and more refined tasks and questions and so on that agents and tools and AI platforms needs to handle in order to get their scoring This new model is focused a lot more on agentic usage and capabilities, and it has shown a very big gap between the leading models to the less advanced models.

46:49Speaker So only 17 models or model variants, I should be more specific, scored more than 50% on this new benchmark or combination of benchmarks. It's a lot of benchmarks actually combined together. So think about how many model variants we have. It's in the thousands, and only 17 of them, crossed the 50% threshold with Gemini 4 Carbon, which is the latest model which we're gonna talk about in a second, scored 69 and leading the pack Another very interesting finding is that now 90% of models perform better on reasoning than on knowledge questions and tasks.

47:23Speaker If you think about just a year ago, it was probably the other way around, definitely two years ago when we did not have reasoning models at all. So the concepts that AI is just a knowledge repository that knows how to pull answers at the right time is definitely no longer true. AI can actually think, reason, and come up with new suggestions and solutions better than it can cite knowledge that exists in the world right now, and that is true across the board on 90% of models If you want a specific example, Claude Opus 5.5, which is the second highest ranking model after the new Gemini model, it scores 78% on the reasoning questions, but only 58% on knowledge, so a pretty big spread The other very

48:10Speaker clear thing in this new way of measuring is that the open weights model trail the leading proprietary models by a much bigger spread than on the previous way of measuring, over a 30-point spread. So Qimi K 3.1 Pro, which is another model we're gonna talk about in a minute, is the top of the open source models scoring at 37%, followed by Minimax 3.1 Flash at 28%, and Qimi K3, the previous model, at 27%, way below, uh, the marks that we see from the proprietary top leading models And now, as I mentioned, we have a new model or kind of maybe soon we'll have a new model from Google.

48:50Speaker So DeepMind has announced that they have a new model called Gemini 4 Carbon. Uh, this comes just two weeks after their previous, Gemini model that they've released This model, as I mentioned, tops the new artificial analysis index, so it is now, on that index, the best performing model on the planet.

49:11Speaker It is has state-of-the-art in creating code, in cybersecurity, uh, defense. It is scoring very, very high on all these different benchmarks, so including 88.2% on the SEC Bench Pro for vulnerability detection, 54.8% on Exploit Gym, for and 71.0 on the CWE Bench for patching. So again, it is not just good at writing code, it is very, very good at protecting the world from cybersecurity attacks, which going back to our initial point in this episode, is a very good thing.

49:43Speaker It is extremely good at advanced, complex, long-horizon capabilities, uh, that should allow the agents to run for very long periods of time while still being effective in what they're doing It also shows very high levels of alignment, and in some of the scores, it achieves 0.0% successful cheating rate on like Terminal Bench 4 and stuff like that.

50:04Speaker So very well-aligned model with really advanced capabilities, extremely successful model from Google. And as I mentioned, this comes only two weeks after the release of their previous model. That means that they had the initial training run on this model for a very long time, and now they're iterating on top of the base model that they've created.

50:22Speaker And this new base model seems to be extremely powerful and Sundar Pichai said that they're expecting to see the iterations happen even faster. I don't know how much faster it can go than two weeks from a release of one good model to a great model, but that's what they're planning to do. So expect to see even more from Google.

50:39Speaker So Google is definitely back in the game. Not just that they're back in the game, they're at the top of the game right now, and they're planning to be even better and better, faster and faster than we've seen before. This means for all of us that we have a lot more to choose from. So if for a while it seemed that we have a two-way race between Anthropic and OpenAI, now we have xAI with really solid models.

51:03Speaker We have Google with right now the top of the line model, and Meta with Muse with a very powerful model. So we have a five-way race, a real race. That means for all of us, we're gonna get better models cheaper, but it also means that we are spiraling out of control faster because there's a bigger, stronger competition on who's gonna get there first.

51:25Speaker God knows where that thing that we're going to is Now Google also made another announcement this week. They released Genie-4, which is a new real-time multimodal world simulation Now, in addition to its ability to do everything that Genie-3 did much better, much faster, much cheaper, it is also the first real-time model, like world model, that generates audio, including ambient sound, music, and interactive dialogue alongside the actual simulation.

51:53Speaker This means that you can create a fully interactive universe of whatever kind you want, including the sound in it, in one path in real time Now, the other thing that this model introduces is the ability to control the simulation in real time with natural language, such as changing the weather, moving the cameras, introducing new objects or people, all in real time, all with natural language.

52:16Speaker World models are a very interesting kind of models, and there's several people in the industry, the most vocal about this is Yann LeCun, that are saying that without world models, we will never achieve real AGI. this is making very big steps in that direction that you can replicate, generate, or anticipate any kind of universe you want in real time without actually building anything up front.

52:37Speaker What does that mean for the gaming industry? I'm not exactly sure, but it doesn't sound very optimistic to me to the gaming industry. If I can just request whatever the game needs to look like, what kind of, scenario I wanna have, what kind of graphics I wanna have, what kind of interactions I wanna have, and it will generate it for me in real time, and when I get bored, I can ask it to change it, so it looks a little different, operates a little different, or becomes harder or easier depending on what I feel, doesn't sound like something that's gonna impact well companies who invest tens of millions of dollars and months to develop similar things right now.

53:06Speaker And as I mentioned earlier when we talked about the new benchmark, we have a new Kimi model that is scoring very high. So Kimi K 3.1 Pro is now the leading open source model on some of the things it is scoring on specific benchmarks as good or better than GPT-6 Astra and Meta Muse Wonder And on the artificial analysis, uh, general index, the one from before, scores 11 points higher than its predecessor, Kimi K3 I'm not gonna dive into each and every one of the benchmarks, but it is doing extremely well approaching or touching, again, the top models from the West, so like Opus 5.5, and Astra 6 and so on.

53:43Speaker As I mentioned on the new test, which is supposed to separate kinda like real work from benchmark work, it is not scoring as good as the closed source proprietary models, but it is right now the most advanced open source model in the world, and it is definitely good enough for probably 95% of the work that you're doing, if not 100% of it OpenAI also released another model.

54:08Speaker It is 6.1 Terra, which is a farther, faster, smarter, and significantly cheaper AI model than everything they released before. So for context, if you don't know, OpenAI has four different levels of models. They have Astra, which is their top of the line, then Sol, which is their largest of the regular three models, and then they have Terra in the middle, and Luna, which is the smallest one.

54:30Speaker So it goes star, planet, moon. Basically, that's the order other than Astra that sits on top of that. So they've released the middle model on a new version of it that is doing extremely well. It is standard API prices are 50% less compared to GPT-5.6 Tera. So if you're comparing it to the previous model, it is more capable, faster, and costs half Like everything else today, it is geared for more agentic use with much better computer use performance And it is built to reduce hallucinations and error rates, and it's a very big improvement from the previous models Now, staying on the topic of OpenAI and cost, OpenAI is apparently luring away Anthropic's key enterprise customers with significant discounts,

55:15Speaker sometimes as high as 90%, and it seems to be working for them, at least according to this article on The Information Jonathan Cornelsen, uh, who is the CEO of DataCamp, uh, reported that OpenAI's new pricing strategy slashes companies' AI cost by 80 to 90% compared to what it is right now Some accounts are talking about 40% that has been reduced, uh, and so on.

55:39Speaker But either way, that is a very big savings for companies So this is definitely a very aggressive discount mechanism for OpenAI to grab market share away from the very successful enterprise, competitor, Anthropic. How long can they keep that? That's a very big question. What we have seen recently is that the cost efficiency of the OpenAI models is significantly better than the Anthropic ones, and so it'll be very interesting to see how Anthropic responds to that.

56:10Speaker That being said, with all of that in mind, where does that put these two companies when it comes to their IPO? if they can sell their stuff so cheap or if their only way to grab clients is to give discounts, can these be sustained? So there are a lot of bigger questions as far as the AI bubble and the IPO costs and so on and so forth.

56:28Speaker But as of right now, OpenAI is investing huge efforts and huge discounts in order to win customers away from Anthropic Staying on OpenAI and their strategy, they have made an interesting shift to also start buying TPUs from Google and not just GPUs from, NVIDIA. so far, OpenAI has been a major, major customer of NVIDIA, probably their largest one.

56:50Speaker And now they're going to start buying TPUs from Google, which is Google's tensor processing, uh, units, which is the chips in which Google is running their capabilities on. Now, they're obviously not moving away from NVIDIA. They're still very close partners with NVIDIA, and they committed crazy amounts to buy GPUs from NVIDIA, but they're going to expand beyond that, to buy Google TPUs.

57:12Speaker And Google TPUs business is gaining significant traction with Oracle also starting to buy it as well. so it is definitely not a one-horse race at this point when it comes to selling more and more compute. A lot of it has to do with the fact that just not enough of it, and everybody's buying whatever they can from whoever can give them such chips.

57:30Speaker Now speaking of spending on AI infrastructure and what does that do to companies, Microsoft apparently is offering a voluntary buyout to 60,000 employees in order to offset the spending pressures of the AI infrastructure So if you wanna understand what 60,000 means, first of all, it's 60,000 people that are gonna be potentially without a job if they take these exit packages, which again, are voluntary at this point, at least.

58:00Speaker This is one quarter of their global workforce Now, these severance structures are, at least for US employees, 60 days of base pay, during which an employee's pay stay on the payroll, plus up to 39 weeks of base pay based on seniority and tenure, uh, beyond the time they stay on the payroll. And there are different levels of employees in Microsoft that get different levels of potential severance compensation.

58:24Speaker But this is an insane number. Again, this is a quarter of the global workforce of a huge employer that is being nicely pushed out of the company for now. Nobody knows what's gonna happen once this, buyout is going to expire, which is at the end of this year Now, if you've been listening to this podcast, you know what I think about this.

58:47Speaker I think on one hand, they have huge investments in infrastructure that they have to somehow counter. But on the other hand, they are generating software. They are writing code. They are servicing clients. All these things can be done with AI at a completely different scale than it was possible before, and they just don't need so many people.

59:07Speaker But letting go of 25% of your employees is a very, very significant move, and it shows what I said all the time. I think we're going to see very high levels of unemployment like we've never seen before with people who previously made very significant salaries. So developers and other workers in Microsoft do not make 30 to $40,000 a year.

59:28Speaker No offense to anybody who makes that kind of money, but we're talking about people who are making over 100,000 and in many cases over 2 and $300,000 per year that are gonna be left without a job and left without a pay once the severance pay package expires. And whether they will find a new job or not is depend on how much other companies are gonna follow the same path.

59:48Speaker And if they are software companies, they are most likely going to follow the same path There is a big story in China this week where the government is now pushing companies to not use AI and not to let people go and actually retrain people for different things. so far it is just a suggestion from the government, but as we know from the Chinese government, suggestions are, let's say, recommended if you don't want to get in trouble, and then later on they become an actual law.

1:00:15Speaker And I have a feeling that we will need to start seeing these kind of laws everywhere. I'm not saying it's a good or bad thing. I'm saying it's necessary because if you get to the point where you have so many people unemployed, it doesn't matter how cheap you can make the services. There's nobody who can actually pay for these services or products, and that means your economy collapses.

1:00:37Speaker So the fact that the Chinese governments are now doing active things to prevent that means they're seeing real signs of that, and I may make this into a full deep dive section next week to talk about this because I think it is important enough That's it for this week. I hope you found this episode and this podcast in general very helpful and valuable.

1:00:57Speaker If you have, please share it with more people. So click on the share button right now, and since you're opening your phone anyways and looking at it, if you're not driving, then you can also find the link to our course that starts this Monday, as I mentioned, on November 2nd, the multi-agent orchestration course.

1:01:14Speaker So if you wanna learn how to build really advanced solutions and be able to use them next year, beginning of 2027, consider this an early New Year's resolution to automate more and more things in your business. Come and join us. There's a few seats left. We'll be back on Tuesday with another how-to episode.

1:01:30Speaker And until then, have an awesome rest of your weekend.