AI's $6 Billion Problem (And Google's Monetization Headache)

Binance says hackers used fake job interviews and an internal AI tool to pull off the $6.3 billion heist

Alex Shannon: I keep going back and forth on this — I think I actually land on the side that this is a good thing. Like, in a twisted way, this story needed to happen.

Sam Hinton: Really? Because I read the same story and I came out the other end deeply uncomfortable. Six point three billion dollars, Alex. That’s not a wake-up call, that’s an alarm that’s been going off for years and nobody answered it.

Alex Shannon: No, I hear you, but my point is — if this is what it takes for every organization to look at their internal AI tools and their hiring pipelines and say ‘we have a catastrophic vulnerability here,’ then maybe the lesson is worth the pain.

Sam Hinton: That’s a very generous framing for the largest crypto theft in history. ‘The lesson was worth it.’ Tell that to whoever lost money in this thing.

Alex Shannon: I’m not saying it isn’t horrifying. I’m saying the method — fake job interviews combined with an internal AI tool — that combination is something every company needs to understand right now. We’ll get into the full story.

Sam Hinton: Yeah, we will. And I promise you, the more you hear, the less philosophical you’ll feel about it.

Alex Shannon: You’re listening to Build By AI, I’m Alex Shannon — and yes, we are already fired up and we haven’t even gotten started yet.

Sam Hinton: And I’m Sam Hinton. We’ve got a massive episode today — Google’s monetization problem, a startup predicting human behavior, Canva laying off over a thousand people, and that Binance heist we just teased. It’s a lot.

Alex Shannon: And this is actually a big moment for the show — starting today, Build By AI is moving to one bigger weekly episode every Wednesday. So instead of daily drops, we’re going deep on the whole week’s AI news in one place. Same great coverage, more depth, every Wednesday.

Sam Hinton: Think of it as the AI news show you actually have time to listen to. Everything that mattered, nothing that didn’t. Let’s get into it.

Google hasn’t figured out how to make AI pay like its search engine

Alex Shannon: Alright, let’s start with Google, because this story is one of those that sounds like a niche business story but is actually a massive signal for the whole industry. So here’s what happened — Alphabet reported that search queries are growing at their fastest rate in five years. Five years. That’s genuinely impressive.

Alex Shannon: But here’s the tension: Google’s Chief Business Officer Philipp Schindler came out and basically acknowledged that the company has not figured out how to make AI as profitable as its traditional search engine. So you’ve got more people searching than ever — but the AI layer on top of that is not generating revenue at the same rate.

Sam Hinton: Yeah, and this is the core paradox that Google is living in right now. Search is their golden goose. The whole business model is — user searches, user sees ads, advertiser pays Google. It’s elegant, it’s been printing money for two decades. AI overviews? The model gives you the answer directly. You don’t click. You don’t see as many ads. The pipeline breaks.

Alex Shannon: So is this an existential threat to their ad business, or is it more of a ‘we need to figure out the new model’ kind of moment? Because those feel very different.

Sam Hinton: It’s closer to the second, but the pressure is real. Think about it this way — Google built the best highway in the world, and now they’re putting a helicopter service on top of it that takes people directly to their destination without driving the highway. The highway still exists. People are still using it more than ever, apparently. But the toll booths on the helicopter route haven’t been built yet.

Alex Shannon: That’s a really interesting way to frame it. And I think what makes this tricky is that Google is essentially competing with itself. The better their AI gets at answering questions, the less users need to click on the results where the ads live.

Sam Hinton: Exactly. And here’s what I think people are missing in this conversation — the query growth number is actually a bit of a double-edged sword. Yes, more queries is good for data, good for training, good for the platform overall. But if each query is now being answered by an AI overview that short-circuits the ad click, the revenue per query is going down even as volume goes up.

Alex Shannon: OK but let me push back a little bit. Google is not a company that’s going to sit still on this. They’ve got Gemini, they’ve got subscriptions, they’ve got enterprise products. Is there a world where they just route around this problem entirely?

Sam Hinton: Absolutely there is. And that’s probably the bet they’re making. If you can convert some percentage of heavy AI users into paying subscribers — Google One AI Premium, Gemini Advanced, whatever the product is — you’ve got a recurring revenue stream that’s actually more stable than ad revenue. The question is whether they can convert at scale.

Sam Hinton: But here’s my honest take: this story is a sign that even the most resourced AI company in the world is figuring this out in real time. Nobody has cracked the AI monetization code the way search cracked the advertising code. That’s not just a Google problem. That’s an industry-wide open question.

Alex Shannon: And for businesses watching this — especially if you’re in the content or advertising space — this matters a lot. If Google is restructuring how search works around AI, your whole SEO and paid search strategy from two years ago may not be the right playbook anymore.

Sam Hinton: Keep an eye on what Google announces around ad formats in AI search over the next two quarters. That’s where the rubber meets the road. When they figure out how to put ads in AI overviews in a way that doesn’t feel terrible, the whole picture changes.

Binance says hackers used fake job interviews and an internal AI tool to pull off the $6.3 billion heist

Alex Shannon: OK, we have to get into the Binance story because this one has been rattling around in my head all week. So according to early reports — and this is from a single source right now so we’re treating it with appropriate hedging — Binance disclosed that hackers pulled off a six point three billion dollar cryptocurrency theft. That is, if confirmed, the largest crypto heist in history.

Alex Shannon: And here’s the method, which is what makes this story genuinely different from your average breach: they used fake job interviews to get access, and they exploited an internal AI tool to actually execute the attack. It wasn’t brute force. It was social engineering combined with AI exploitation.

Sam Hinton: And this is where I want to pick up from our cold open, because Alex, your ‘maybe this needed to happen’ take — I understand the logic, I just think it understates how alarming the specific combination is. Fake job interviews as an attack vector isn’t new. We’ve seen North Korea use that playbook before. But bolting an internal AI tool exploitation onto that? That’s a step change.

Alex Shannon: Walk me through why the AI tool piece is so significant. Because on the surface, you might think — an AI tool is just software, software can always be compromised. What makes this different?

Sam Hinton: So think about what internal AI tools are designed to do. They’re built to be helpful. They’re given access to data, to systems, to workflows — because that’s how they provide value. The attack surface of an AI tool that has permissions inside your organization is dramatically larger than a traditional piece of software. If you can manipulate the tool — either by getting access to it through a compromised employee or by prompt injection or some other method — you’re not just exploiting one vulnerability, you’re potentially accessing everything that tool has permissions for.

Sam Hinton: And the job interview piece is clever because it’s the human layer. You get a person inside who then has legitimate credentials, and suddenly the AI tool’s access is their access. It’s not hacking the AI — it’s using the AI as the lever once you’re in.

Alex Shannon: That is genuinely unsettling. And it makes me think about every company that has deployed internal AI tools with broad access to systems — which at this point is a lot of companies. How many of them have actually audited what those tools can touch and what the blast radius is if they’re compromised?

Sam Hinton: Honestly? Not enough. And I think there’s been this assumption in a lot of AI deployment conversations that the security review happens at the model level — ‘is the AI safe, is it aligned’ — but the security review for access permissions, for what the tool can actually do inside your systems? That’s often treated as a secondary concern. This story, if it’s confirmed, makes that inexcusable.

Alex Shannon: OK, so I want to give a practical takeaway here because I think this is actionable for people. If you’re deploying AI tools inside your organization — whether you’re a small team or a large enterprise — what should you actually be doing differently right now?

Sam Hinton: Three things. First, audit what your internal AI tools have access to. Apply the principle of least privilege — the tool should have access to exactly what it needs and nothing more. Second, treat your hiring pipeline as a security perimeter. That sounds weird but it’s real — if social engineering through a fake candidate can get someone inside with credentials, your onboarding security matters as much as your firewall. Third, and this is the big one: have a response plan specifically for AI tool compromise. Most incident response plans weren’t written with this scenario in mind.

Alex Shannon: Really practical stuff. And again, this is based on early reports from a single source, so we’ll be watching for confirmation and more details. But the method described, if accurate, is a blueprint that other bad actors will study. This is not going to be the last time we see this kind of attack.

Sam Hinton: No, it is not. And the fact that it allegedly resulted in six point three billion dollars makes it the kind of case study that gets analyzed for years. Keep an eye on the technical disclosure from Binance, because how much detail they share will matter a lot for the broader security community.

Anthropic is working on a system to prove AI models’ identity

Alex Shannon: Alright, let’s pivot to something that actually connects really directly to that last story, and I love when that happens. Anthropic — Claude’s maker — is developing a system called FlashProve. And the idea is pretty simple to describe but actually kind of profound in its implications.

Alex Shannon: FlashProve is designed to verify that AI responses actually come from the AI model they claim to come from. So if someone tells you ‘this response was generated by Claude’ — FlashProve gives you a way to verify that. It’s like a digital signature, but for AI output. This has been verified by multiple sources, so we’re confident in the core facts here.

Sam Hinton: OK so when I first read this, my immediate reaction was — wait, why does this need to exist? And then I thought about it for about ten seconds and the answer became very obvious. Because right now, there is no way to verify that an AI response came from where it claims to come from. I could take a Claude response, strip the metadata, re-label it as GPT-5, and you’d have no way to know the difference. Or I could build something that claims to be Claude but isn’t — a spoofed model — and use it in a product.

Alex Shannon: And the Binance connection here is interesting, right? Because one of the attack surfaces in AI exploitation is exactly this — impersonation. If you can make a malicious model impersonate a trusted internal tool, you’ve got a very powerful attack vector. FlashProve is essentially building the infrastructure to prevent that.

Sam Hinton: Totally. And I think there are actually three separate use cases here that are worth unpacking. The first is enterprise trust — companies deploying AI want to know that the model they’re calling is actually the model they contracted for. The second is regulatory compliance — if you’re in healthcare or finance and you have to document which AI made a recommendation, you need cryptographic proof, not just a label. And the third is exactly what you said — security. Identity verification for AI is the same problem we solved for websites with SSL certificates.

Alex Shannon: That’s a great analogy. Because before HTTPS, you couldn’t verify that the website you were on was actually the bank’s website. Now that’s a solved problem — your browser does it automatically. Is FlashProve essentially trying to be the HTTPS of AI?

Sam Hinton: That’s exactly what it’s trying to be. And here’s where I want to push back on myself a little — I think this is genuinely valuable and Anthropic deserves credit for working on it. But I do have a concern, which is: who controls the verification infrastructure? If Anthropic builds FlashProve and it becomes the standard, they have a privileged position in verifying which AI responses are legitimate. That’s a lot of power for one company.

Alex Shannon: That’s a really valid point. It’s kind of like — imagine if only one company could issue SSL certificates. That would be a problem.

Sam Hinton: Right, and we actually had moments where that was almost the case in the SSL world and it caused problems. So the ideal outcome here is that FlashProve sparks a conversation about open standards for AI identity verification — where you have multiple providers, interoperability, and nobody owns the certificate authority monopoly. Whether Anthropic builds it as an open standard or keeps it proprietary — that’s the question I want answered.

Alex Shannon: For developers and companies building on top of AI models right now — this is worth paying close attention to. If AI identity verification becomes a standard requirement for enterprise contracts or regulation, being early to adopt that will matter.

Sam Hinton: And for regular users — think about what it means to know that the AI response you’re reading is genuinely from the model it claims to be from. That’s a trust layer the internet currently doesn’t have for AI, and FlashProve might be the first brick in that foundation.

Google will let schools disable its AI tools, but not without sacrificing Search

Alex Shannon: Let’s talk about Google and schools, because this one — I want to be honest — kind of annoyed me when I first read it, and I want to unpack whether that’s fair. So here’s the situation: Google announced that schools can disable its AI tools. Parents, administrators, educators who are worried about kids interacting with AI — OK, you can turn it off. That sounds reasonable.

Alex Shannon: Here’s the catch, confirmed by multiple sources: if you disable the AI tools, you also lose access to Google Search. Because AI features are now so baked into Google Search that you can’t have one without the other. So schools are being handed a choice that isn’t really a choice — accept AI in your classrooms, or lose the most used search engine on the planet.

Sam Hinton: Yeah, ‘annoyed’ is a polite word for how I feel about this. Because I understand the product integration logic — from Google’s perspective, AI is Search now, they’re not separate things. But for schools who have genuine, thoughtful, well-reasoned concerns about AI in education — age-appropriate use, academic integrity, critical thinking development — this is basically Google saying ‘sorry, it’s a bundle deal.’

Alex Shannon: Is there an argument on Google’s side here, though? Like — is there a world where this is actually defensible?

Sam Hinton: I mean — I can steelman it. The argument is that AI-enhanced search and traditional search are now genuinely the same product. You can’t have a 2024 car without a computer in it — that’s just what cars are now. And maybe Google’s position is that trying to offer a ‘dumb’ version of Search is technically complex and creates a worse product for everyone. I get that argument.

Alex Shannon: But you don’t buy it.

Sam Hinton: I don’t fully buy it, no. Because the power dynamic here is really uncomfortable. Google Search is not optional for most schools — it’s so embedded in everything they do that removing it would be massively disruptive. And Google knows that. So offering a ‘choice’ that is effectively no choice at all — that’s using market power in a way that feels coercive, even if it’s not technically illegal.

Alex Shannon: And this connects to a broader question about what happens when AI gets integrated into foundational infrastructure. It’s not just search. We’re going to see this with productivity tools, with communication platforms, with everything. The AI gets baked in and suddenly opting out means opting out of the whole thing.

Sam Hinton: Exactly. And I think the education sector is a really important canary here, because schools have populations that include kids who can’t consent for themselves, whose parents have strong feelings, and where the stakes of getting AI interaction wrong feel high. If we can’t get the right policy in place for schools, we’re going to have a much harder time getting it right for other sensitive contexts.

Alex Shannon: So what should schools actually do if they’re in this position right now? Because ‘complain about Google’ doesn’t feel like a strategy.

Sam Hinton: Honestly? The leverage is collective. Individual schools don’t have bargaining power with Google. But school districts, state education departments, national education bodies — if they coordinate and push back as a bloc, that changes the conversation. And I think this is actually a moment for regulators to pay attention too. Because the question of whether companies can bundle AI features with essential services in ways that eliminate meaningful opt-out — that’s a policy question that somebody needs to answer.

Alex Shannon: Keep an eye on how education advocacy groups respond to this, and whether we see any state-level legislative action around AI opt-outs in schools. This story has legs.

RAPID FIRE

Alex Shannon: Alright, rapid fire time. We’ve got four more stories to get through and I want to give each of them their due even if we’re moving quickly. First up — Fal, an AI inference startup. Early reports suggest Fal has doubled its valuation to sixteen billion dollars after raising one point three billion in a funding round led by Bessemer Venture Partners. If confirmed, that is a staggering number for a five-year-old company. Sam, what’s the signal here?

Sam Hinton: The signal is that inference — which is basically the compute that runs AI models when you use them, not the training but the actual real-time processing — is where the money is right now. Training a model is a one-time cost. Running it at scale for millions of users? That’s an ongoing, enormous, infrastructure problem. If Fal has cracked that efficiently, sixteen billion dollars starts to make a weird kind of sense.

Alex Shannon: So it’s essentially the picks-and-shovels play. Everyone needs inference, so the company selling inference infrastructure wins regardless of which AI models become dominant.

Sam Hinton: Exactly that. And the fact that demand is surging — Bessemer doesn’t put one point three billion into something without serious conviction. Keep an eye on who Fal’s customers are, because that tells you everything about where the AI usage growth is actually concentrated.

Alex Shannon: Next — Trajectory. Verified by multiple sources. A one-year-old startup just raised forty-five million dollars and is now valued at three hundred million. The pitch: making AI model training accessible to non-technical experts. So you don’t need to be a machine learning engineer to train your own AI model. Sam?

Sam Hinton: OK this one really excites me because the gap between ‘I want a custom AI model for my specific business problem’ and ‘I can actually build one’ is still enormous for most organizations. Right now you either hire expensive specialists or use off-the-shelf models that may not fit your use case. Trajectory is trying to close that gap, and if they do it well, the downstream economic impact is huge — suddenly every mid-sized business can have AI that’s actually trained on their own data and processes.

Alex Shannon: It’s the no-code wave applied to AI training. And three hundred million valuation at one year old tells you investors believe the market is real.

Sam Hinton: The question is execution. Making model training genuinely accessible without letting people create something dangerous or just bad — that’s a hard product problem. But the ambition is exactly right.

Alex Shannon: Third rapid fire — and this one got a reaction from me I want to be honest about. Canva. Early reports from The Verge suggest Canva is laying off twenty-two percent of its workforce — that’s twelve hundred and seventy-one employees. CEO Melanie Perkins made the announcement. Twenty-two percent is a substantial cut for a company that was considered one of the more people-first tech companies out there.

Sam Hinton: Yeah, and I want to be careful not to speculate beyond what we know, because this is a single source report and we don’t have full details on why. But what I will say is: when a design platform company — one that is very directly in the crosshairs of generative AI tools — restructures this significantly, it’s hard not to read AI as part of the equation.

Alex Shannon: The design space has been genuinely disrupted by AI. If Canva is building more AI-native features, the workforce mix you need to do that might look really different from the workforce you had two years ago.

Sam Hinton: And twelve hundred people is not a rounding error. That’s a real human toll. Whatever the business rationale, I think it’s worth sitting with the fact that AI-driven restructuring at this scale is going to become more common, and we need better policy and support systems for the people it affects.

Alex Shannon: Last rapid fire — this one is wild. 42AI just came out of stealth with sixty-two and a half million dollars in funding. And their claim — take a breath — is that they’ve built a world model that can predict human behavior. In novel situations. Situations it hasn’t seen before. Sam, go.

Sam Hinton: So I want to be very clear: this is a single source report and we should hedge heavily. If confirmed, the claim is extraordinary. And you know what they say about extraordinary claims. I’m genuinely skeptical of ‘predict human behavior in novel situations’ as a product description — that’s either a profound scientific breakthrough or very ambitious marketing language.

Alex Shannon: It’s the kind of claim that could mean anything from ‘we have a really good recommendation algorithm’ to something genuinely unprecedented. We just don’t know yet.

Sam Hinton: Exactly. But the sixty-two million suggests serious investors took it seriously enough to write a check. I want to see the technical paper. I want to see what ‘predict human behavior’ actually means in practice. Because if it’s real — even partially real — the applications and the ethical questions are both enormous. Watch this space very carefully.

BIGGER PICTURE

Alex Shannon: Alright, let’s zoom out. Because we’ve covered a lot of ground today and I think there’s actually a coherent thread running through almost all of these stories. If you look at everything we talked about — Google’s monetization struggle, the Binance heist, Anthropic’s identity verification, the school AI opt-out problem, Canva’s layoffs, 42AI’s behavior prediction claim — what do you see?

Sam Hinton: I see the same thing in all of them, honestly. It’s a trust and infrastructure problem. We have rapidly deployed AI into critical systems — search, finance, schools, design tools — and the foundational layer of ‘can we trust this, can we verify this, can we control this’ is lagging badly. FlashProve is a direct response to that. The Binance hack is a catastrophic consequence of that gap. Google’s school dilemma is a governance failure caused by that gap moving faster than policy.

Alex Shannon: And the funding stories — Fal at sixteen billion, Trajectory at three hundred million, 42AI with sixty-two million — those tell you that capital is still flowing freely into AI infrastructure and capabilities. The build is not slowing down. Which means the trust and governance layer has to catch up, and it has to catch up fast.

Sam Hinton: Here’s what I want people to really think about: the Google monetization story is usually framed as ‘Google has a revenue problem.’ But I actually think it’s a signal of something bigger. The fact that the most sophisticated advertising machine in human history can’t figure out how to make AI pay the same way suggests that AI fundamentally changes the value exchange. Users get more value. The old extraction model breaks. That’s not a small thing — that’s a structural shift in how digital economies work.

Alex Shannon: So the question going forward is: who figures out the new value exchange first? Is it subscription? Is it enterprise licensing? Is it something we haven’t invented yet? Because whoever cracks that becomes the Google of the AI era.

Sam Hinton: And that company might be one year old right now. It might be Trajectory, it might be something we haven’t heard of yet. That’s what makes this moment genuinely exciting and genuinely terrifying. The map is being redrawn in real time, and everyone — companies, governments, schools, individuals — is navigating with instruments designed for a different world.

Alex Shannon: The question I’m leaving this week with: in a world where AI is embedded in everything and you can’t easily opt out, what does meaningful human agency actually look like? I don’t think we’ve answered that yet. But I think the stories this week are all, in their own way, trying to find the edges of that question.

Sam Hinton: And that’s worth sitting with. Not in a doom-scroll way — in a ‘I’m going to pay attention to how this resolves’ way. Because the decisions being made right now, in boardrooms and classrooms and regulatory offices, are setting defaults that are going to be very hard to change later.

OUTRO

Alex Shannon: That’s going to do it for this week’s Build By AI — our first ever weekly Wednesday edition. It felt good to go deeper on everything rather than trying to sprint through it in fifteen minutes every day.

Sam Hinton: Yeah, I think this format is going to be really good. More time to actually dig in. If you got value from today’s episode, subscribe wherever you get your podcasts — it genuinely helps more people find the show.

Alex Shannon: We’ll be back next Wednesday with everything that matters in AI that week. Until then, keep building thoughtfully — and maybe audit what your internal AI tools have access to. Just a thought.

Sam Hinton: Seriously though. See you next Wednesday.