Googleの可能性、AI資金の蛇口は閉まるのか、PHP脆弱性と「ブラック・オクトーバー」

What is Google's potential? We'll get into it after my conversation with Google CEO Sundar Pichai. Is AI's money spigot finally turning off? And is Black October for cybersecurity as a PHP bug and plenty more end up opening the doors to bad guys in technology? We'll cover it all on a Big Technology Podcast Friday edition right after this. This episode is brought to you by Genesis. What does it actually take to put agentic AI to work across an entire enterprise? I recently spoke with Genesis Chairman and CEO Tony Bates about why orchestration is emerging as a competitive advantage in customer experience and what companies need to get right in the AI era. Then I sat down with Adam Mitchell, head of enterprise business solutions at Voya Financial, to go inside a large-scale agentic CX transformation, including the business decisions, trade-offs, and lessons along the way. You can watch both conversations now on my YouTube channel today. It's never been more challenging to get Frontier Compute.

How do you know where to look, who to trust, or how much to pay? Akash is an open cloud that provides easy access, flexibility, and peace of mind. That's how companies like Razer were able to cut their inference costs by 15 times. Visit akash.network.com for a free consultation. That's akash.network.com for a free consultation. Welcome to Big Technology Podcast Friday. edition where we break down the news in our traditional cool-headed and nuanced format. We have a great show for you today. We're going to talk about Google's potential and cover a little bit of my interview with Google's CEO Sundar Pichai, which was conducted earlier this week. We're also going to talk about whether the money spigot for AI is closing as companies like Meta say that the demand for money is starting to outpace the supply. And then finally, Alex Damos is going to join us and talk about the PHP vulnerability that's turning the tech world on his head. Joining us as always on Friday to do it is Ranjan Roy of Margins. Ranjan, great

to see you. Welcome back. Wednesdays with Sundar, Fridays with Ranjan, just another week at the Big Technology Podcast. As it should be, as it should be. And you know, it was really interesting for me to put some of our long running questions to Sundar, including what matters more, the product or the model. You know, I did tell him, I said, hey, listen, people are arguing that you just need standard models and you can put that into action in your products and that is a good strategy for Google. Sundar wasn't having that at all. He said that Google's committed to the frontier, that it's always been about the frontier. It won't have it any other way. And you can make good products with standard models, but the frontier is basically frontier or bust for Google because the frontier gives you visibility into what's coming. It also helps you improve your models faster. As you develop more models, you can rely on the frontier to help you make them better quicker. All right. We know what team you're on, Sundar. He's on team model. What do you think about that statement from him? Does it make sense? Does it make your thinking about how the product

is more important than the model in today's age of AI? Does it shift that thinking at all? Okay. It does not, at least the first part. The second part I do think is interesting around, like, the more at the frontier you are, the better you can actually train future models, and it starts to give an edge overall. I think there is something to that. I think in terms of like overall progress for a company, that being at the frontier does mean it can move faster. And I think anyone, I mean, the speed and velocity with which Anthropic and OpenAI actually just launched, everyone feels it, everyone sees it. And I think that is kind of validation of that part of it. But in terms of I still more firmly than ever, and I think anyone who uses, and I will say this, Google's suite of AI products and how disjointed they are. And we've talked about this a lot around like Gemini and workspace versus Gemini standalone. Like you can feel that lack of cohesion in the product within the entire Google ecosystem.

And yes, I know it is a big, large, sprawling, massive company. But I think actually a little more focus on the core product itself, Sundar, I'm going to say it, I think would be valuable for Google. Okay, so that is another interesting point that we can bring up about the interview. And by the way, if you're new to the show, if you came in after seeing the conversation, just a bit of housekeeping. We do these news breakdowns, Ranjan and I, every Friday. And then we'll have a flagship interview like the interview with Sundar, typically every Wednesday. So next week, for instance, we're going to have Wharton professor Ethan Malek talking all about the productivity gains or the invisible or non-existent productivity gains. We get into a bit of a debate that we've seen after the AI era has kicked off. So this is kind of how the show works. Really happy to have so many new viewers, and we appreciate you being here and listeners. All right. So Google's product priority. This, to me, was actually one of the more interesting and subtle takeaways from the conversation with Sundar. And we're going to, of course, get into the company's potential very quickly,

but I think it's important to go over some of the highlights. So I asked him, I was basically like, when, I mean, not even basically, this was the question straight up, is when is search within Gmail going to work as good as Gmail being plugged into ChatGPT? Right? Very interesting answer from Sundar. As opposed to saying, yeah, we're working on getting AI into all of our products. You know how he answered? He said this. He says, we're trying to bring this all together in a unified agent, first for enterprise and then coming for consumer. He says, you can address this agent whether you're in workspace search in Gemini. We're putting a lot of time and effort into making sure as we build out our authentic experience that unification and that capability works well across all these services. You know, and then I said, well, that's a good answer to the interface question because I asked him what the interface of the future looks like and he goes, it is. So I think I, and you tell me if you're reading this in a different way, but I thought this was very telling. I almost read this as, you know, in a way Google's going headless, where it's instead of having separate product areas,

which is kind of the way that it operates now, everything is going to be unified through the agent. And the priority is building those integrations within the agent as opposed to building the integrations within the product themselves. That sounds smart to me. Let me turn to you. What do you think about that strategy? I think that's absolutely what they need to do. Again, the idea of having Gemini as like this little button at the top right of your Gmail and your Google Sheets and going and staying in that kind of being constrained to that traditional product layer and AI becomes an add-on to AI is the core UI and you're chatting with it, you're asking it to go through your email. It's the way so much of AI has already moved, and I think it's the right choice. I will say, knowing how siloed a lot of different parts within Google are, and even I think that reflects in the product, it's interesting to me, and actually, we've joked in the past that Google's McKinsey background makes them uniquely suited to recognize that.

Sundars makes them uniquely suited to actually pursue this organizational change that is required at a company like Google. But who would own that product? Is it Google Cloud, which owns Google Workspace? DeepMind is no longer DeepMind, right? No, it is. Google DeepMind. Google DeepMind. Is it the Google search? Like, who actually owns that at a company like Google? To me, that would require some massive internal transformation to actually make that successful. But isn't it pretty simple? Like don't you just create a new team to own it and you then change the way that you evaluate the other products, right? So let's say you have the Gmail product group. You're going to say you still work on Gmail, but as opposed to, let's say, engagement or user numbers being your North Star, is now going to be how many times people repeatedly query Gmail through the agent.

That actually sounds perfect. And I think that's exactly how it should be. Now, come on. It could be tough to swallow because it's a lot of centers of powers within a big company. Centers of massive amounts of powers and hundreds of billions of dollars of revenue. I mean, this is to just say, oh, by the way, your worth is now number of queries rather than you have been evaluated forever and your entire department's worth is based on daily active users and total utilization to, oh, yeah, you're just an API call now or not necessarily an API call, but here's just some queries to this other new shiny product. But again, it's interesting to me because a company like Google, I think more so than – I mean, actually, Microsoft, it's the same thing. And I think we see that disjointed nature across a lot of their product offering. So for the giant hyperscalers, that is the bigger challenge to me.

And actually, as we're saying this even more now, Sundar, team product I'm going to double down on because I think if you can solve that, You have every user in the world sitting and using your product, the vast majority of them, like, make that work well. And just being around the frontier should be good enough. And, you know, you're hitting on an interesting point here, which is that Google really is this company with so much going on and yet many dilemmas. Another dilemma is where to allocate the compute. And you mentioned Google Cloud. And I think there is a well-documented and obvious clear conflict of interest between AI research and cloud compute and cloud services within Google. Let's say you have a bucket of compute. You can give all that compute to cloud services. And because of the demand for AI compute, chips, et cetera, you could make massive profits.

We'll talk about this in a minute, but Google Cloud Platform grew 82% this past quarter. And that's with a lot of the compute capacity going to AI research, right? So if you were to put even more capacity to cloud, I wouldn't be stunned if they overtook Azure and maybe even AWS at some point in cloud computing. But basically, you know, I spoke with Sundar about this, and what he said was, everybody is short compute. He said, today, when you look at everything, everyone is short, not just us, other companies, and everyone wishes you had two times what you planned for because these decisions that you have today in 2026 were made in 2024. And this is probably the toughest part of his job is allocating the compute between cloud profits and research, which is a cost, but can put it in place to be able to lead the AI game. And, of course, as much as we might clamor for the product, it's always like, well, where's Google's latest model? And they finally have one with Argon that's performing quite well on the benchmarks.

I love this because, actually, that decision, where do I allocate the compute, is the ultimate expression of team product versus team model. Are you going to push it to the research side and the actual frontier training side, or are you going to push it to Google Cloud, which is going to be the much more application and product-heavy side of things? Again, it's not pure product, but it's still in the much more kind of like general utilization of that compute. So I like it, Sundar, and he's going to show us which team he's on based on where that compute gets allocated. And it sounded like from the interview that it's still going to be leaning towards the research side. Is that a fair understanding? I think basically what he was trying to say was research is going to eat. Here's the quote. Our cloud business is at any given moment in time foregoing short-term revenues because we are making compute trade-offs to support frontier AI development. So very clearly, cloud is sacrificing money.

And then he said something else, which I thought was interesting, and I don't think the world has picked up on. So what he talked about was that some constraints are good. Even our AI researchers have some constraints, which are also good. That's what inspires creativity. That's what helps you come up with a more efficient algorithmic breakthrough. If you were to run that logic by, let's say, the AI researchers at Anthropic or OpenAI, I think they'd have a heart attack. So I think it is very interesting to hear his perspective that, like, yeah, I'm going to give you the amount of compute I think is necessary to advance the frontier, and we're committed to the frontier, but some constraints might be good for you. That was interesting. That is a very different philosophical shift that no one picked up on yet. No, it's great to call that out because I actually think that is very relevant. Again, to go back and forth, though, if he's saying on one side the entire industry is short, compute, and my biggest challenge is allocating that compute.

But I do like this idea because, again, there's a lot of kind of talk around how it was actually a lack of access to abundant compute that actually pushed Chinese innovation more on the algorithmic side. We've seen innovations with decision models and JEV recently and actually doing more with less, certainly that kind of constraint forces innovation versus you have infinite access to compute and that's going to be how you approach all these problems. I think that is interesting that he's saying that. again what does that actually mean for his what do you think those meetings look like do you think they have like a meeting where all the most powerful people at google sit and then wait to be told what which pile of compute they're going to get access to it's like one chart goes up and it's like all right sorry cloud research is uh research gets this many i do think that there's a version

of that. I mean, I'm not privy to the details of the meetings. We're just hypothesizing. This is fanfic. Google fanfic. Right. But I mean, it would not be stunning for me to think that there are people with projects within the company who are like, I'm going to need this much compute or I'm leaving. And those are the decisions that he's faced with. I'm sure that happens. I mean, Noam Shazir, for instance, who's been in and out of Google and found a character AI in the middle, he left. He went to OpenAI. Why do you think that happened? I mean, I could just imagine the conversation. And they spent a lot of money bringing him back. Yeah, yeah. Could have just invested in some, well, I guess they couldn't have just invested in some more GPUs. Right. Hence the short compute, but. Well, they're making TPUs. The problem is that the decision takes two years in terms of build-out, probably even longer, honestly. But, yeah. And then on the cloud side, it's probably just like, you know, I could see the cloud. I could see Thomas Corian making an argument.

Listen, we are in a rare position where after lagging as the number three cloud computing company, cloud computing service, we can make a legitimate push for, you know, at least number two and potentially number one. And I mentioned this before. Google Cloud Platform, almost an 82% growth rate this past quarter, year over year. Their revenue last quarter was $24 billion, $24.8 billion. Azure revenue last quarter was billion AWS revenue was billion AWS 37 growth Azure, 42% growth. Google Cloud, 82%. Sorry, Azure, 42% and Cloud, 82%. They have a meaningful shot here at at least passing Azure. I don't think people are recognizing this. If these numbers continue to pace for maybe another year or two, they're a solid number too. And if you're running the cloud business, man, it's like you make that pitch.

And I mean, it's a company that's been struggling to build a cloud business for a long time or really clawing at it. This is the moment to capitalize. But of course, it comes with the expense of research. What is your decision as a hypothetical Sundar? Flood the cloud, man. Flood the cloud. Cloud's got to eat. No, I don't know. I don't know. I wouldn't want to be in that job because it's so hard. How about you? I think, again, I love it. It comes back to do you believe the product or the model is going to be the – You're giving the cloud the money. I'm giving the cloud the money. I'm giving the cloud the money. I'm telling my researchers do more with less, come up with innovations that actually make this more compute efficient. And, again, they have the distribution base. They have the – like every whatever billions of people will be interacting with their products. that to me, I just more so than ever, most people don't need to be at the frontier.

Most people do not need to be on the frontier model for most tasks. Not all of us are doing, not everyone needs to be doing 3D modeling and whatever else. Like it should be fine for them to focus on creating the most, again, like the instincts of the world. We talked about this with Muse. No one even cares what model it's actually running on at any given time for a given task. All everyone cares about is that they are magical product experiences. And to me, if Google, where's your agent? Where's the personal agent? No, I highly disagree with you on this one. I think that if the unified agent is going to be the product experience that people are going to be competing about, having the smartest and most capable one is going to matter. Because right now what you've seen is Muse. I mean, just imagine you built Muse, but the model underneath it was 10 times smarter. You would obviously use that one. Would you? What does that actually mean? Because Opus 5.5 a lot of the times does not give me what I need it to give me.

It like overthinks. It overdoes things. For simple things, it's not necessarily the right answer. Well, so this is what I'm trying to say. is like with model innovation, you won't be saying it doesn't give me what it needs to give me. You'll be saying, remember we talked a lot about how important it is that a model understands the intent of the query and routes it the right way. And so if you have one that's incredible at that, that thinks just the right amount for when you have a sort of low intensity task and a lot when you have a high intensity task and you can trust that it gets it right, and it has the EQ to know when it should interrupt you or not, that I believe will be the winning product. I mean, otherwise, what are you competing on? Your mascot. Your computer. Certainly the mascot. And what would Google's be? Let's start naming Google's personal agent. Or is it just Gemini? Don't let Alexander Wang get in. Leave Gemini alone. Leave Gemini alone. Do not sexualize Gemini.

That's my platform. I'm voting for that. Yeah, I still disagree. Again, I think like the majority of use cases and ideas that like people will be interacting with to constantly get that product layer better and better and better at answering it. Even again, right now with Muse, like when should it respond quickly? When should it wait? Even tiny little things. I think Instinct, one of the things I love it does, is a little emoji reaction just to show you, like, it registered your query. Even it might take a little longer to actually come back to you. Like, there's so many ways people can actually, these products can try to help guide people along the way versus just sheer model intelligence. Are you still using dots? I am using dots. I'm liking it. I think that we should definitely make time to talk about just the potential of Google because every time I speak with Google, it's clear to me that if they're able to get all their ducks in a row, so to speak,

they are so well positioned, probably better than anybody else. So think about it. I wrote a little list about Google's potential if it hits everything right. So it could have an agent. Give me a list. The list is this. It could have an agent that works seamlessly with its services, right? Which, by the way, since so many people have 13 products with a billion users, that agent can be extremely valuable if it's working across your Gmail and calendar natively. It can have an AI for information retrieval, right, which is search. It can license out its models via the API, so make a licensing fee there. It can resell or self-compute and potentially even chips via Google Cloud platform. It has Waymo. And then, of course, it has tens of billion dollars either invested or committed to be invested in Anthropic. So it will have a huge chunk of that company. Name me a company outside of Google that has the ability to hit on all of these fronts,

from unified agent to model licensing, autonomous driving, and then, of course, the investment. I can't think of one. No, I completely agree. And I kind of love this list. And but again, to me, this is like what we have seen to date is all of those potential channels also present potential complexity or not existing complexity that actually make things difficult for them to execute. Again, why when I ask maps in Google Maps, do I get a completely different answer that's worse than when I ask Gemini the same question? Why did Google Maps integration? Again, I'm asking like very pedestrian things that should be solved. And that's where each one of those, I mean, when you said it, 13 products, each with a billion users, they know everything about all of us. They should be able to serve us all better. The entire search, like better search experience they should own.

Shopping. Actually, you didn't even have shopping here. I actually think they're incredibly well positioned in agentic commerce to actually make a dent on Amazon. But that presents the challenges, too. Who owns what? Who? What power centers are you kind of fighting against? So to me, and again, Sundar, if he can figure this one out, the organizational structure that gets them to the greatest potential on this, if anyone can do it, he can. And so I don't disagree. Okay. That's a good way to wrap the segment. We have 10 minutes before Stamos comes, and I want to talk to you about this concept that the AI money spigot may be running out. Remember that these companies have been funding AI infrastructure through VC, through their own free cash flows. I'm talking about OpenAI and Anthropoc, but also the public companies like Google and Amazon and Meta. And then the free cash flow ran out. And so now what do they have to do?

They have to get creative. They're using debt. We've talked in the past about off-balance sheet debt. They're doing equity sales to be able to fund this, including Meta, which just split its stock and took away even more voting rights. So it's able to sell that stock, I think, and invest it in building compute. The market's had it. Meta reported these earnings, and it's down 9% on the week. and this is from the Wall Street Journal. A regulatory filing from Meta noted that recent significant capital raises from peers like Alphabet and SpaceX alongside anticipated fundraising from Anthropic and OpenAI have created an increasingly challenging fundraising environment. Oh, as a result, the company would more proactively raise funds before competitors absorb the market's available capital. This is the first time we're hearing, honestly, of the capital potentially running out for companies like MetaGoogle, OpenAI, and Anthropics.

So there is a limit here to the amount of money that can be raised. And so, Ranja, my question to you is compute has been the constraint all along. Basically, it's been as quickly as compute can come online. It will be bought and used, but the bottleneck has been the amount of compute. that you can build. Is the new bottleneck the money? Because if Meta is running into these problems, what is the potential for this entire AI build-up to get stopped in its tracks by the simplest thing of all, the amount of cash you can have to build it out? Well, I think we talked about this last week, and one thing that I'm actually kind of happy about is we're actually entering a period where there's rationality being applied to any kind of capital request, capital raise. Like when there is SpaceX this week trying to raise $40 billion of debt and

actually seeing that the actual yields they would need to pay were much higher than expected. Even in Meta, I think they're supposed to raise, the capital commitments now have risen to $825 billion up from 700 billion dollars in july like the market is saying okay it's going to cost you more to get that money which is good it should just from like a capital is finite anyways and i do think there's but but it's gonna we're gonna see more of that then you have the whole neocloud story with core weave and others and then i mean obviously there's like a great deal of potential turmoil there. But I think, to me, is the reason capital is becoming more constrained because people don't buy the compute story or that they've seen that it's actually very difficult to build these data centers and that it's not going to happen as quickly as possible? And going back to Sundar's point as well, do you think the market as a whole has said actually this idea that the entire world

will be short of compute is actually validated? Or do you think people are just like, actually, I'm not going to give you money to fund data centers because they're hard to build? I believe it's neither of these things. Okay. So I have two other reasons. One is the investors are starting to do the math on what it's going to take to make these products profitable. This is a little item from a Wall Street Journal story that I read, which actually did the math on what Muse would need to generate to pay back some of these investments that Facebook is making. Remember, Facebook raised its guidance from $700 billion that it's going to spend on AI infrastructure to $825 billion. This is the amount of money they anticipate Muse building in one of the good scenarios. If Muse skyrockets to 100 million users next year paying $20 a month, it would only add $24 billion to Meta's annual revenue, about 12% of next year's anticipated capital spending. associated advertising fees could add more to the pop but probably not a ton meta made about

67 dollars of revenue per daily active user from at mostly from ads um from in the year and year to end of the third quarter for 100 million users that would add another 6.7 billion again this is a company that's spending uh 100 billion those are 825 billion on uh on ai on the ai build out So basically, in order to make this work, the company needs to get Muse to a billion users. And even then, we have questions. I think trying to come up with what is the scaled business model of personal agents like Muse, and normally I'm very skeptical and people try to avoid that question. For a company like Meta, I still think it's still a bit early. Again, trying to apply the would you pay $20 a month, Are you trying to extrapolate out the existing ad revenue per daily active user? I still think we don't know what the business model around personal agents is going to be,

or will there be a sustainable one? But I mean, agreed. If the market starts asking those questions, it's a huge problem for a lot of companies, especially startups, because I think to date, especially on personal agents, The assumption is whoever just starts getting people draining tokens and using it nonstop and getting completely habitual to it, the money will come. So if people start asking that, that is a problem. So that's one for me is people are actually doing the math here and they're like, does this work? And the other side of it is I think they're fatigued. The system itself has put so much money into these companies and it's still experimental. And the belief was there, but eventually that hits a wall. As for the Financial Times, OpenAI has cut its valuation target for its latest funding round after facing resistance to a $1.4 trillion price tag from some investors. It's going to be $1.25 trillion now. Deeper in the story, this was the rationale.

Multiple people with knowledge of the Morgan Stanley-led process said demand from clients was weaker than in the past. There has been zero clamoring, said one of the people. All the prior times OpenAI have been on, there's been a frenzy. This time it's a bit of a thud. Some investors were also suffering from fatigue or indigestion after having invested heavily in prior rounds. OpenAI raised $40 billion in March 2025, $122 billion in March this year, and Anthropic has raised close to $100 billion this year. And that might be the most simple and obvious explanation here. It's fatigue. And I am happy if this is the case. You know what, everybody? I hope you all have a little bit of indigestion. Let's all just take a moment. Even the fact that there was zero clamoring, that in the past there was much clamoring, is a reminder that everyone has been way too kind of intense about it. Because, again, the fact that we're calling it a thud when the jump in valuation is going to be almost 50% higher than its last funding round is crazy.

Like the extrapolation, the expectations is still, I mean, just maniacal. So the idea that like that is a thud, I think it's healthy for the markets overall. So if everyone can take a breath, pause to start moving forward again, I am happy. So do you think that this means OpenAI briefly? Do you think this means OpenAI is no longer on track to a successful IPO? or is it just sort of, I mean, usually these things, once these things stop going up, they start going down in the real way? I think it all depends on Anthropic and what happens. I mean, it is kind of crazy to think about. Like you're sitting at, you're invested in OpenAI heavily. All that matters is how the market receives Anthropic. Any kind of like, let's like try to evaluate the financials on a comparable basis, none of that is going to matter. If Anthropik gets hit there no way OpenAI is going out Ronja before we move over to Alex Damos I just got to ask you one more question Like you know the financial markets you know the systems where you know

people raise money and people invest it, having worked, again, on trading floors in the past. Is it possible money just ran out? Like, did they just take all the available money? Well, I mean, if you look, it is UAE investors and CO2 that were coming in on this. So still, we talked about where do the deepest pockets live. I mean, it's like the sovereign wealth funds are the final boss, basically. Final boss. Yeah, and we've reached them. Short of, you know what? Where's the next government treasuries? That's your next, your last bastion. Yeah, no, I've been saying for a long time that the Gulf states are the final boss of fundraising before IPO and opening up. Here we are. That's the game. All right. Look, when we come back, we're going to have a great discussion. I think people are largely underappreciating how intense the PHP vulnerability has been, and along with a number of other security vulnerabilities that have emerged this month, people are calling it Black October for cybersecurity.

security. And if that doesn't make sense, and there are still things that I am trying to learn about it, good news is we are going to unpack it with someone who really does understand this world and is able to do it in a way that's accessible and hopefully help us wrap our heads around how big of these vulnerabilities, how big these vulnerabilities are that AI is finding in systems like PHP and what it means because we talked with Alex basically as the hugging face hack happened. He said it was a big deal then. It obviously proved to be a huge deal. He and I just texted this morning and I was like, I want to talk to you about PHP. And he goes, not just PHP. It's been an absolutely insane couple of weeks. So we'll speak with Alex about the implications of the insane couple of weeks and what the ripple effects are going to be when we come back right after this. This episode is brought to you by AvePoint. Everyone's racing to roll out AI right now. Co-pilots, chatbots, agents doing real work.

But here's the part nobody loves talking about. All that AI runs on your data. And most teams have no single way to see it, secure it, and prove it's under control. That's exactly what AvePoint does. For 25 years, they've been the trusted layer beneath the world's most demanding data, now extended across your entire AI estate. Your data, your cloud, and the agents acting on your behalf. It's how more than 28,000 organizations deploy AI with confidence. So innovation scales without scaling risk. Avpoint, the unifying trust layer for AI. Head to avpoint.com to see how enterprises deploy AI with confidence. Learn more at avpt.co slash bigtechnologypodcast. That's A-V-P-T dot co slash Big Technology Podcast. And we're back here on Big Technology Podcast Friday edition. We have, of course, Ron John Roy of Margins with us, as always on Friday. And today we bring in our guest for when the proverbial excrement hits the fan.

Alex Stamos is here. He's the Chief Information Security Officer at Cognition. New job, same Alex. Alex, great to see you. Welcome back to the show. Yeah, thanks for having me. I'm glad that I am the Grim Reaper of this podcast. The last time you were here on the Friday show, we were talking about the hugging face attack. Basically, right after it happened, before the report was released, and you're like, guys, you don't realize how bad this is. We certainly didn't. Ranjan was still calling it marketing. then it turns out that then it turned out that hundreds of bots had autonomously decided to hack this hugging face company in a fairly unprecedented attack that has basically driven the AI anxieties through the roof Ronson I'll give you a second to defend yourself here the severity of the situation I will recognize I definitely have some questions on like the actual how this vulnerability was found exploited

And what's that mean? Because I'm still going to push back that they autonomously went out on their own and started doing nefarious things. Like definitely trying to understand, is this kind of like targeted cybersecurity hacking? Is this people who know they want to exploit things and are doing that? And then we'll get into it. We'll get into it. I was wrong that it was just marketing. I throw my hands out. Alex, you're here for a seminal moment here on the show. So Alex was spot on last time. And now, you know, as we're talking, you know, Alex's belief is that this, what we're about to talk to you about is the next big thing. And as I mentioned before the break, I was like, I got it. We got to talk to you about PHP. And he said, it's not just PHP. This past few weeks have been insane. So I'll just read the story. And then, Alex, you could take it away. This is the headline from the Wall Street Journal. AI just found a bug that puts up to 70% of the world's websites at risk. The creators of PHP, a technology that powers as much as 70% of the world's websites, disclosed a critical security issue that could be leveraged by hackers to take control of any system running its software.

That would allow an attacker to steal information or launch a cyber attack on anyone visiting a hacked website. Good times, 70% of the web, not a big deal. So Alex, just talk to us a little bit about how this was found. It was found with AI and what this has led to and what it presages for what comes next. Yeah. First, it's at the stage. We'll talk about PHP first. But this has been about the craziest 30 days or so in the history of our industry of cybersecurity. And so PHP is what's catching the headlines. But for people to do this professionally, there's actually lots of stuff that's happened just the last couple weeks. So the PHP bug specifically, like you said, PHP is a very popular language for writing websites easily. So very few kind of enterprise quality websites are written in PHP. The interesting counterexample is Facebook, although Facebook doesn't actually use real PHP. They use a typed version of PHP called Hack.

But other than that, it is very hard to think of like a big company that uses PHP for a banking website or a big application or something. That being said, it's extremely popular for companies to make side projects, right? Here's my marketing site. Here's a smaller application. It's extremely popular in a number of frameworks and hosters like WordPress that hosts lots of sites. And so it is still very widespread. The bug was found by a project that was being run by OpenAI called Hack the Planet, in which they have been providing their... Great name. Yeah, exactly. Right. That is marketing. I'll agree with Ronjan on that one. But there's been, as we talked about actually on stage at your conference, there's a lot of controversy about the access to the best cyber capabilities from the two Big L labs, OpenAI and Anthropic, who at differing levels restrict the ability for people to use their software

to do vulnerability research. And so as a response to that, both companies run these programs where approved people have access to use their models to go find bugs. This is part of this where they were actually giving this away to free to bug finders, a company called Calif, which is a kind of excellent security consultancy that does a lot of AI enabled vulnerability research. They used it to find the PHP bug. And then that was actually several months ago. And there's been a coordinated effort by a bunch of companies to try to get ahead of this issue of what's called coordinated disclosure. What kind of off the rails in the last couple days, however, before the release, in that whenever you do these coordinated disclosures, you start, obviously, with a small circle, and it gets bigger and bigger and bigger, and it started to leak out, and we now have evidence that bad guys started exploiting this well before the public release. There was a controversial dump by the CEO of Chain Guard, which is a company that will make their own patches and push them to customers.

But looking back, he might have been in the right. It's a little bit controversial. There's a big discussion. Anyway, the bug is a big deal because it is a pre-authentication remote code execution. So it means you do not have to be part of the web app, right? So there's been a bunch of web app vulnerabilities where you need to be logged in or you need to be accessing some kind of functionality. This just works on the front end. And it's remote code execution. It is not an incredibly easy exploit to write. It turns out to be quite a complicated issue. But people have figured it out, probably with AI. So that means that I just visit the site and my stuff can be jacked. What it means is as a bad guy, you can just connect up and attack the system without having a login and without having special access. So even if the PHP site, the front page says, log in here, you can't access anything else, you're still vulnerable. And you can't fix the bug without patching PHP, which is under your web code. They only released a patch for the latest family of PHP,

and there's a lot of PHP sites on older versions. And so upgrading that breaks your site. So it's not just as easy as going in and saying, you know, run your update program, whatever your operating system is. For a lot of companies, they have to actually refactor their code to be compatible with the latest version of PHP 8 because it was not backported. There turns out to be some, like, unofficial backports in patches that you can get, but those are not as widely available. So it's a big deal. What does that mean in terms of the vulnerability of every website that has been built with PHP in terms of the hacker can just take it down, they can access actual underlying databases, they have full access to whatever is in their whole stack? They can take over the computer at the same level as the web server is running, which by definition has access to all of your data, right? So what has been happening since just a little bit before the announcement is when you have a bug like this, you have a bunch of different bad guys go out

and try to break into as much stuff as possible, and then they'll figure out later what they actually care about. And so what we've been seeing is automated exploitation of hundreds of millions of attempts to break into millions and millions of websites that are going on continuously. So these will be Russian ransomware actors who will like to take these things over, install malware, and then figure out who they want to actually ransom. It is probably what is called the Chinese cottage access industry. So there are – within China, there's effectively like an eBay for hacking. And so there's providers in China that will break into a huge number of sites. We've seen this with other bugs. And then they will later sort through and then figure out who they want to sell access to inside of the PRC. There is probably the Iranians who have – Our eBay is vintage sweaters, and their eBay is ransomware in your consumers. Yeah. It's not quite like a full eBay site from my understanding, but, like, there's basically an ecosystem here. And it's quite smart because in other countries like Russia, you have this problem where if there's a desire to hack something,

everybody rushes in, and they all step on each other's toes. So the PRC has the ability to kind of de-dupe taskings for their private access industry. So, anyway, I expect that those people, we have seen this in the history. There was a very famous SharePoint bug in Microsoft where these Chinese access vendors broke into something like 80,000 SharePoint sites that were on the Internet and then sold them later to Chinese companies. And then these guys break in. They plant malware. And then sometimes what they do is they actually fix the bug so somebody else can do it. What we have seen in the past is you'll have them fighting over control of these systems. So whoever gets there first, part of one of the things they have to do is close the door behind them. Otherwise, other people will break in, remove the malware from the first hacker, and then patch it themselves. And so that is effectively what's going on right now is you've got dozens or hundreds of hackers, of hacking groups, attacking tens of millions of websites and trying for control. And then in the long term, what you'll see is they'll then have their list of this is what I control.

And they'll search through it of like, oh, there's a government website. This is a hospital. Like, what is the interesting stuff that now that I have control of it and I have malware there, I'm going to actually steal data or do a ransomware? So by control, just so I understand it correctly, it's you control as users are entering information into the website. You're receiving all of that. But does it also mean you then have access to anything else like in the back end or it's both? You will have access to what is happening live, but also probably everything in the database. Yeah. And then it also gives you a foothold inside the organization. So you're now behind the firewall. You're in their production network. And so you've got your first foothold inside of their network. You can use that to maybe break in deeper. So this is just the start for the PHP bug. This is the initial access vector. There will be some of these attackers are probably breaking in and then just dumping the entire database out so that they can just use the data

to try these actors, what they're doing. Yeah, yeah. It's just like you just copy it all and then you figure out later what was interesting. 95% of it's going to be useless. 5% of it will be gold. And then some of them will say, okay, great, now that I have a foothold inside this company, I can use this to scan the rest of their systems, find other vulnerabilities, and go out of back. The other thing that I don't know, I don't have any good statistics on where we are, but in some cases they're also using AI to do the attack, which means the traditional I have to wait doesn't really exist because you can just have your AI agent go do the rest of those steps automatically. That, from what I've heard, is not as widespread, but that is going on in other situations that we can talk about. For example, a Chinese group broke into a bunch of South Korean banks with completely automated AI attack chains. And so in this case, if you have that attack system set up, you could use this to pivot inside of the network to go to different systems. So I'm just going to give some stats that Matthew Prince from Cloudflare gave.

In the 10 hours before it was publicly announced, this PHP vulnerability, across the Cloudflare network, we saw thousands of attempts to exploit it. That jumps to tens of thousands in the 45 minutes post-announcement, hundreds of thousands within two hours, and millions now. So this is certainly leading to a ton of hacking attempts. Just explain it for us. If you're not cybersecurity-brained, what do you think the next couple months are going to look like? Obviously, the damage is being done as we speak, but this is one of those things that it'll be latent until it starts to emerge. So what do you think the next month, two months, three months, six months are going to look like as a result of what we just saw? First off, it is now basically impossible to get an instant response company to respond to. So anybody qualified was already booked before this. So you'll have a bunch of – Yeah, what's an instant response? Instant response. So like a forensics firm to help you do this work, they were already basically completely booked because of everything else going on before this. So if you are a big company but you don't have a team that can do your own instant response, you're in real trouble here because it will be very hard to get a Mandiant or CrowdStrike or SentinelOne or any of the other companies that do this work.

to respond quickly. So first off, you have lots of people trying to either do their own forensics or to recover by themselves. We will see the breach certifications are going to start happening for the companies that actually have good security teams because they will hopefully have detected that this happened and then they will be able to do their own forensics But that is going to be spread over a significant amount of time because there going to be a lot of people who patch it and don't realize that within the 48 hours it took them to patch, somebody got a foothold. And so then we'll see the secondary breaches over the next months as attackers now use that foothold to go deeper and deeper inside these companies. for you know ransomware actors they want to get paid so in those cases they will contact those companies and ask for money for the government actors we won't hear anything unless uh there is some kind of interesting research that finds their attack servers or something like that and so what you'll see is a huge amount of data is leaving these companies a huge amount of customer data we're also gonna see a huge problem if everybody's passwords just got stolen right so uh the the

Number one way any individual human is hacked is that they use the same password on multiple websites. And what we're about to see is in the dark webs where these are sold, you're going to see these dumps of tens of millions of passwords being sold over and over and over again. And for a disturbingly large number of people, that password is going to work from their email. Their bank. Their bank, yeah. Especially if you get access to somebody's email, their life is over because then you can use that to reset almost all their other passwords. And so we're going to see a ton of kind of individualized fraud, spam, that kind of stuff is all going to hit a new high over the next six months. What do we do as everyday simple folks who are just logging into our bank accounts and emails? And what do we do? I mean, use one password. What do you recommend? Yeah, so for normal folks, there's nothing you can do to actually stop the breaches. So what you need to do is you need to make sure you're using a different password on every single website.

The easiest way to do that is a password manager. One password is one I like. There are other ones as well. So you need to install a password manager. And now if you've got that, you know, the name of your cat, 123, was your password, then every site that you're using that on, you need to now change it and use a randomized one. My Pebbles 123 password is now exposed. There you go. Goodness gracious. So you've got to do that. You should just be very – I think what's starting is we've lived through this era where people just kind of assume things were secure. That was never true, but it kind of worked because the number of attackers that were active was – Was finite, basically. Right, was finite. And so if you just kind of blissfully walk down the street through a bad neighborhood, you just got lucky over and over again. Well, the neighborhood's gone really bad. We've gone from like, you know, a bad neighborhood in L.A. to like, you know, Somalia, 1990s, right?

Like it is completely – the threat profile has completely changed. What is the likelihood something bad is going to happen to you if you are walking down the street? And so I think people are just going to have to be much more paranoid. Obviously, there's a password thing. They're going to have to be much more paranoid about what websites they use. Like people just throw their personal information everywhere, login everywhere, and you're just going to have to – You can't really do what we call vendor risk management as a normal person of asking them lots of questions. But you can get a feel of like, is this fly-by-night forum the kind of thing I want to put my personal information in? And I think what even normal people will start using stock puppets, right? So people will start creating secondary Gmail accounts for all the crap stuff they do, and then they'll have their important account for their bank and all that. And you're going to want that secondary thing to be as disconnected from your real identity as possible. And I'd be looking out for frauds and scams, right? the number of spam is going to go way up. You're going to get way more texts and emails from people you know asking you to click something. People are going to be getting phone calls because all this directed graph information

about who knows whom and their phone numbers have all been leaked. And so you're going to get a lot more of the bad texts and the phone calls of like that are all AI driven now of, you know, your grandchild's been kidnapped and all that. So like people are just going to have to be much more paranoid going forward. And we talked about this, but I think unfortunately, this is probably the new normal for several years where the pace of crazy things that have happened in the last 30 days might be this might be up here. But we're not going back to the baseline. It might drop a little bit, but it is possible we are now at a heightened level of kind of operational tempo for bad guys and for defenders for at least several years. And briefly, what's it going to mean for corporations and governments? So the consumer level we got down for the big guys, what's going to happen to them? Yeah, so on this I'd like to expand it a little bit because the PHP bug is a big deal. But just to make it clear, basically within 24 hours of the PHP bug, there was a large number of Node flaws.

So Node.js is also a popular web application server framework and is much more likely to be used by enterprises. Now, fortunately, the bugs there are more complicated to exploit, but they are significant. We've had a bunch of breaches over the last couple weeks. There was a certificate authority that had their certificates stolen, so kind of a not-so-great certificate authority. These are the people who the little lock icon in your browser, they're the ones that make that work. And if somebody steals those certificates, they can create certificates. They can create an Amazon certificate or a Google certificate that is also trusted. I'm seeing this list. It's, yeah, it's PHP, Node.js. This is from X, so you can tell us if one of these are wrong. Java, GitLab, Railway, Signal, WordPress, LiteLLM. They're calling it Black October. Black October, yes. It is spicy. I mean, this is what we talk about on stage. Like, this is the start of the takeoff of what the new normal is going to be based upon just the number of vulnerabilities that are being found,

but then also the operational tempo that attackers are able to stay at going forward. All those breaches you're talking about, so like Railway, Fireworks, those are almost certainly not PHP. So that is just attackers being good at their jobs and that being enhanced by AI. And finding API keys for everything you built on Railway, which I did insecurely myself on some side projects in the past, so I went back and, yeah, yeah. Same with Fireworks. So there's a lot of attacks against infrastructure ingredient companies where you see, and that is why I expect those attacks are professionals. That is not like a 19-year-old. That is probably somebody working for a commercial, like a commercial vendor who's doing state-level work because folks are breaking into people who are important in the AI ecosystem, neoclouds, hosting providers. And then once you have those API keys, you can turn around and go back to the cloud. So, yes, our enterprises. So, enterprises are dealing with PHP right now.

They're dealing with Node. They are asking all of their vendors what they're doing, right? So, I have to – I keep on having to fill out these – send these emails saying we do not have PHP at this company. But we did have Node. Everybody does, but we fixed it as quickly as possible. what you know the companies are like kind of panicking for their vendors for themselves what this is reminding them is they probably have a lot of stuff they didn't know about like a bunch of these PHP sites are like somebody set up a website for a conference they held two years ago and it's still up yeah exactly and if you architected things correctly then if you pop that nothing bad happens for too many organizations Once you're there, you're in a DMZ with a bunch of other servers that you can now compromise. Or there was a bunch of data there that now you have to do forensics because you had the registration lists for everybody who went to your conference. That just got downloaded.

You're going to be ransomed for that. Or it was downloaded for a state actor who's going to use it for future whatever use. And now you're going to have to do a breach notification. So there's a bunch of companies that are realizing that their attack surface is actually much larger than they expected because of all of this random crap, a lot of which is actually in the cloud. that makes it hard for them to pull together and say... And governments, like I saw the NSA? The NSA was talking about this? Yeah, I mean, NSA published, CISA published something. So, yes, it is a big enough bug that you have the U.S. government. Unfortunately, the number of people who do defensive cyber in the U.S. government has gone significantly down in the last couple of years because CISA was kind of decimated by political appointees who don't like the idea of CISA doing election security stuff. And this is showing now the outcome of that, which is we really need competent people in the government to help coordinate response to things like this.

Okay. Even though I am absolutely terrified right now, I'm still going to ask one level above that I'm wondering about. So everything we've described so far is still, again, purposeful, coordinated attacks to exploit leveraging AI to increase the velocity of these attacks. I still – then there's still the entire world of – and then there's reports on banking, like unsupervised agents potentially acting autonomously. Obviously, the hugging face thing, the scandal seemed to be more around agents going rogue as opposed to people controlling them in nefarious ways. It's like, is that yet another layer we need to worry about? Or do we have enough to worry about yet? Yeah, that's just, I mean, so those are two totally valid but completely separate issues. You've got the agents escaping during evals and training. That is going to be, I'm not going to say solved, but the industry has done a lot to reduce that. A lot of things you've been hearing in announcements are stuff that happened in the same time frame as Huggy the Face.

It's just it's all coming out now. Still, it's all kind of leaking out. Uh, you know, maybe that will happen every once in a while, but there's been a lot of changes in the way people do these evals and training. That is not like an actual, I think that is not really an actual, like OpenAI's models going crazy and hacking you is not actually the threat that people should worry about. Um, the other side is, yes, we are people using agents and then being like, hey, you're giving instinct or muse. here's my entire 1Password, go handle my finances for me, go file my taxes, is totally insane. I'm hanging out in shame. I did it. But people are doing that. People are absolutely doing that. And so that's what the banks are talking about, is that they now see all of these interactions with them that are authenticated correctly, so they're not stolen credentials, but are clearly bots doing interaction. And one of the reasons Instinct has a ridiculous valuation from my perspective, but one of the things they did that made them popular is they actually figured out ways to defeat a bunch of the automated bot detection mechanisms from these companies.

And that's why Instinct is useful, because it will effectively bypass the safety protections that are in place at these companies to keep you from doing exactly what you're doing, which is to have your bot interact with the most important parts of your life. so that is you have this crazy tension where you have like the banks and big tech companies and financial institutions and folks who are like I am going to do everything I can to prevent automated access to my systems on behalf of customers and then you have this big demand from customers of I want my life to be easy so I'm just going to turn over all my logins and personal data to a company that's existed for six months I just want to end with this there was a Nikita Beard tweet who he said former head of product at X. He said, there's nothing to worry about. Nothing important is run on PHP or old software except the water and electrical grid, a few government services like our missile defense systems, and a handful of electrified fences at the zoo, specifically for the lions, tigers, and the cobras. So, folks, if you do see a lion walking in your neighborhood this weekend,

just understand that the origin was OpenAI's Hack the World program. Yeah. Hack the planet. Hack the planet. Hack the planet. The planet. Which is a reference to That hackers movie from the 90s I don't know There's another reference I can't make these references that work anymore Because now I work with a bunch of 25 to 30 year olds And I can't make any 90s references Literally I had I mean there's like multiple employees at Cognition Who were students of mine at Stanford And one of them called me Professor And it was like the scene from Indiana Jones Where he comes to Crypt Keeper immediately Yeah But I can make it with you guys you guys are closer in age to me. I remember Angelina Jolie was in that. Yeah, I remember. That was like what computers were supposed to be when hackers were, yeah. Yeah, and they're hacking a mainframe. Yeah, yeah, yeah. There's a lot of neon. It's pretty fun. It's a fun movie. A lot of people watch it, like, in parties at Black Hat Def Con. But, okay, to OpenAI's credit, like, the truth is,

is this bug has existed for a decade. A bunch of the bugs that Anthropic and OpenAI are either fixing themselves directly or are facilitating the fixes of are issues that have existed for decades that have always been waiting there. And so there are people, I just want to push back on this, because there's people on X who are, like, blaming Anthropic and OpenAI. And the truth is these bugs were there, and it is very hard to fix an open source product like this that is so widely distributed. This was handled as well as it possibly could be. It was always going to be bad. And if OpenAI and Anthropic were not doing this work, these bugs would be discovered. So I've got my M5 Studio Max finally came. I'm running an obliterated quantized version of GLM 5.3. And wow, like this stuff, I feel like a much younger man with the reverse engineering I can do, the exploits I can write, the things I can do now on a Mac Studio that leaves absolutely no logs. And so these bugs were going to be found by the open-weight models, the Chinese models. Somebody was going to find them.

We have to fix them. And the only way through this of the two years or three years or five years of dark ages we're going to go through is to find bugs and fix them. And then also we're just going to rewrite a huge amount of software. That is we can't have these applications written in like type and memory safe on languages by humans. We're going to have to refactor the whole thing. And the good thing here is that would have been absolutely impossible five years ago. Now it is totally plumbable, right? Like that's actually a big part of our business at Cognition is helping big companies refactor stuff. And there's all these examples of like some old SAP-based software that was rewritten in eight days and such. And so you could do that, but it takes money. It takes time. It takes access to the best models. It takes kind of a willingness on behalf of the entire open source community to throw away the old stuff and to replace it with something new. And so I think that is one of the things that might come out finally. This is something I've been calling for actually for a while. It's like we just need to rewrite a huge amount of the software that underlies the Internet.

Maybe finally there will be the billions of dollars necessary in both money and tokens thrown into a pool here so that that can get done. Right. Last week, my biggest worry was the honeypot. So it turns out I was not anywhere prepared for where we would be. They're good old-fashioned human honeypots, actually. We're going to look back nostalgically. Exactly. Exactly. Yeah, well, those are hard to copy with AI. For now. Alex, it's always great to speak with you. You know, you've been here with us through thick and thin through the year. And as the stuff keeps hitting the fan, we'd love to keep speaking with you. So thanks again for coming on. Yeah, I feel like this won't be the last time we chat. Well, unfortunately, we'll be back. Oh, God. All right, everybody. Well, have a very comforting weekend. Thank you, Alex, and thank you, Ranjan. We'll see you next time on Big Technology Podcast. Stay safe.