Anthropic’s Backlash, Wasted AI Spend & the New Security Crisis

Hey everybody, happy Friday and welcome to the Techstrong gang. We got a light crew today because, well, it's summer and people are on vacation. But I got to say, just because people are on vacation, it's still getting weird out there, and there's just a lot of things happening that don't seem to quite line up with the way we think or expect them to.

So we're going to dive right in. John, welcome to the show. How are you?

I'm good. This is crazy, right? Alan, I think he wrote a column where he referred to it as the world turns in terms of AI.

Yeah. And it's upside top, bottom, left is right, yes is no. It's just crazy.

Nobody knows what's coming next. Fred, always good to see you, my friend. How are you?

Doing really well, brother. Good to see you, too. All right.

Well, let's dive in, and there's no end of places to start this soap opera, but we've talked about just about everybody else this week, so let's have a chat about Anthropic. It's their turn. It's just getting harder and harder to understand what it is that these guys are after and what their strategy overall is going to be.

On the one hand, sometimes they like to position themselves as the good alternative to OpenAI, and yet, we hear them ditching open source models. We hear them talking smack about a couple of different things here and there, and there's not a lot of love lost for them, either. It seems like maybe it's just a question of degrees.

And now, of course, they're saying, "Hey, don't forget about us. " And I guess they felt the need to come clean on that, or is it just part of their marketing campaign? John, we're all confused.

What's going on with these guys? I'm confused, too. I think of Anthropic as this interesting, it's kind of growing pains in a sense.

Everyone's taking their turn being the piñata in Silicon Valley in terms of AI, in terms of being flogged. It was OpenAI for a long time. Now it seems to be Anthropic's turn, and in a sense, Anthropic is kind of a victim of its own messaging.

I think they've moved from this transition from the safety-focused research lab that does good, like the Google of its era, to this commercial enterprise that's causing this severe backlash among their partners and developers, especially their software partners. So, for instance, let's mention a couple of things. They launched something called Claude Design, which is basically competing directly with the API agents or clients like Figma.

Anthropic has abruptly discontinued older models and raised prices, which has led to accusations that it is guilty of product cannibalization and abandoning its original thoughts and ethos. These strategies have led to startups reevaluating their reliance on this single provider and maybe looking towards multi-model architectures or open weight alternatives. So what's interesting to me is on the regulatory front, there are critics, and I think it is a legitimate argument that Anthropic is using its high-profile advocacy for AI safety and strict frameworks as a kind of a strategic shield to build competitive moats around its technology.

And critics also mention Anthropic's opposition to open weight models and model distillation to stifle open source competition. This was heightened by Anthropic's refusal to sign a joint industry letter supporting open source models alongside Microsoft, Meta, and OpenAI. And I also think what has happened in the Valley is companies each take their turn, as I said earlier, of being categorized as the bad guy for a limited period of time.

So they all take their turns, but I think this is their time in the spotlight, and it's going to be a rough time because for Anthropic, they really want to push this IPO as fast as possible. And I think all this noise really has created this narrative of negativity that they are going to have a hard time digging out of for the next several months. Fred, you work with a lot of these models.

What's your take on what's going on with Anthropic as it compares to any of these things? And is all this maybe inside baseball, and the average developer may not even care? I guarantee the average developer cares, for a million reasons, but I think it's difficult to disambiguate rhetoric from actual facts and behaviors.

And I think some of the challenges are there's a lot of hypocrisy in here about who participates in what and whether or not models breaking out of a sandbox is actually models compromising companies and so on. And the thing that's concerning too isn't whether or not-- We would expect this, right? There are two heavyweights basically slugging it out at the moment.

Both of them have an equitable stake. There are a bunch of peripheral players like, I'll call Meta a peripheral player in this particular case, where the vast majority of the forward-leaning security improvements and also detriments are being caused by the same folks. And does it make sense that this happens, I don't know, a month before Black Hat?

Possibly. Right? Does it make sense- Mm-hmm ...

that there's also an industry-level illustration of disclosure from Hugging Face? Probably. But the challenge is motives and intentions from actions and behaviors, and it's difficult to figure that out.

So I'm with you guys. It's getting weird And I think part of the challenge that people are trying to figure out, back to your question about developers, is look, it doesn't even matter whether it's this model or that model, right? We're now talking about harnesses.

We're now talking about what happens when models that are meant to have guardrails don't perform as they should, agents that don't perform as they should or do things they ought not do. And that characterization is super illustrative of where we are between adoption and the ability to have the same level of insight as we have for all of our other infrastructure. You know one thing I was going to mention, Mike, I'm glad the way Fred termed this is interesting because I'm thinking, yes, was it yesterday?

It all blurs together for me with these stories. We talked about Meta. Remember that?

And Mark Zuckerberg's campaign, his active campaign, to tell us that they want to be the responsible, they want to make super intelligence available to everyone. They don't want to box out anyone because they're playing catch up. And in a sense, I think of this also as kind of a problem that pertains to people like me in the media, where we create these rags to riches to rags storylines.

So Meta was the bad guy, they were the rogue, and now, in a sense, Zuckerberg, through his editorial in the Wall Street Journal, his interviews with the FT and the New York Times, is trying to spin a new narrative that we buy into because we loved to have access to. This access journalism leads to people controlling the story narrative for a limited period of time. So I actually think this is kind of this evolution of a blur between who is good and who is bad for a certain amount of time, while all these companies are working with each other in a certain sense or competing with each other.

We've got all these cross-pollinization of partnerships and spending, and we don't know who is doing what with whom. It's dizzying, but this is what happens when Al always mentions $8 trillion, right? Its potential market.

This is what happens. So is this not also maybe a story that's as old as time? " And so basically, they're sitting there going, "Well, should I really trust these guys?

" And so, Fred, is this any different than the conversation ISVs used to have with Microsoft and the operating system people? I don't think so. I agree with you.

It's the same business harmonics here, and in the same sense, the same people get caught in the lurch, which is the consumers. And so, hey, appreciate you basically saying these models are not going to be available any longer. These are the ones that I've been using.

Now I've got to retest, I've got to spend more tokens to figure that out. I don't have a real understanding of what the value per token looks like per model as this. Then you've got a debate about which weights actually help me get where I need to go, open weights versus none.

And as a person trying to adopt but also drive value in a deterministic way, it's really hard to discern what the correct path is forward. Couple that with things that we don't have in that old paradigm, which is like choosing the wrong vendor may kill you, right? Metaphorically speaking.

And without the level of understanding of the sophistication of that problem or the observability to understand it from the normal acuity you would have, we don't turn stuff on in a network we don't understand until we do. And that's probably the same set of recourses but hasn't matured yet so that we know who actually we need to hold to account here. It's mind-boggling, but there is so much money at stake here that it's an obvious thing that I would expect to see another loop on that cycle.

And then I feel like all this security stuff is just adding insult to injury. And basically, the insult is you're my partner, but by the way, I'm crawling all over your software to look for vulnerabilities with or without your permission and as I'm testing all this stuff. And what am I doing with that information, and what happens to that later on, and am I using that to reverse engineer something that might be interesting or competitive?

Because, well, trust is a commodity that seems scarce. John, what do you say? Yeah.

That trust was one of the first words I thought of when I saw this story. This is just days after the OpenAI security issues, and it feeds into this whole narrative about-- And I'm going to have to give Zuckerberg credit. I'm going to go back to him because he just always seems to know when to swoop in and try to take advantage of a situation, where he's talking about all the doomsday sayers at Anthropic and OpenAI, and then conveniently, these things happen.

Maybe he had some insight, but it's just like, I don't know as a consumer, I don't know who to trust at this point, right? I'm reading these accounts. I know these guys are all embedded with the government.

There is no regulatory oversight. Only Europe seems to be taking any action. And I look at these companies and I just think if I'm a consumer, I'm looking, I'm like, "Oh, well, it broke out of a sandbox.

What the hell does that mean? Does this mean I can't trust it? " Seriously, I think there is a bit of hyperbole that's both positive and negative about what's going on at this juncture in time.

And then throw on top of all of this is the CapEx question we've been talking about and the inordinate amount of money that's being thrown into us, which creates or sparks this fear that there's some sort of bubble percolating, not immediately, but in the perhaps near future, six months, nine months. Mm-hmm. And when you're looking at the results every quarter and looking to see what revenue's coming in and what's going out.

They claim they're generating revenue, and they are. It's just not enough to justify- It's not enough to cover the cost ... all the investment.

Yeah. But, the other side of this too is this is happening across the ISVs and the partners, but if you're the end customer and let's say you're a bank or you're a construction company and you're building software on these platforms, Fred, are you not having the same concerns and the same conversations about if they can do this to Hugging Face, who can they? Who else are they going to do this to?

Absolutely. And it's also the lack of understanding of the requirements of what it means in either direction. So imagine the rest of the sheep are sort of trying to get themselves into a position where they can utilize some of these tools effectively, and now there's a whole bunch of seemingly magical ways that agents can roll through a configured proxy to then get to Hugging Face, find RCEs, then exploit them, and then chain them.

And then that's too much, right? I think that's just too much for the normal person. And what I hear right around the neighborhood when I talk about this is like, "AI's bad.

" And I'm like- Wow ... " But the reality is when we start looking at where the squeeze happens, it's always going to be the consumer. And so your point about, let's say, Bedrock, for example, is a really good one, Mike.

I think one of the risks that we run here, too, is that in my infrastructure, I have access to data, I have immutable proof and truths, I have provenance, I have these other things that help me understand exactly what happened. As we look at ISVs, providers, and even directly these frontier labs and others, we don't have all that. So I think there is going to be a shift.

A bubble will be, okay, look, if I have no other options, I might have to figure out how to do this myself. And I think that's inherently good for the consumer, but it is a steep climb from where we are today. Mm-hmm.

So is there a way to think about this where, hats off to OpenAI, Anthropic, they invest billions, and they have some of the leading-edge models. But as an ISV, do I want to have maybe a series of cascading gates that basically say, let me minimize my risk with these guys, and then I'm going to default to maybe the thing that's the least expensive and will do the task as I need it without maybe creating a dependency on somebody that I don't fully trust? Is that becoming a factor in my strategy?

What do you say, Fred? I have to imagine that is, because the exchange on revenue for that would dictate that, let alone all of the other components of that, which would be security. My infrastructure security based on these models, also something for me to consider.

My expense on security and controls to support that same endeavor also have to be counterbalanced. And then there's the notion of what happens when you customize models. So why wouldn't I take advantage of that given these are the things that are happening.

And commodity models. Not everybody can afford the luxury of grabbing the latest five, six model to go do this from OpenAI or the latest Opus model or what have you, Fable. Those are expensive.

They are not expensive in a way that most people understand either, which I think we'll talk about later. But the real risk is translating that to the implications it has, both for you just gathering information, building applications, but also what some of these large provider models are doing now, which is some of the security work. The Hugging Face example illustrates both of those cases, where it's adversary using LLM-based agentic workflows.

Plus, you also have the defensive part of that, which is how Hugging Face found that. So it's interesting to think about the cost models associated with that. I'm curious if that becomes a little bit more like the mutually assured destruction model that we've seen a few years ago, maybe coming back.

Maybe cool. So John, is the financial press starting to figure this out? Because to your earlier point, they do engage in a lot of what I call, quote-unquote, hero journalism, where they interview the CEO, and the CEO is the driving force of everything that happens at that company, and it's the great man story.

And history has shown the great man story tends to be b******t, but are they- Oh, I've been exploited on that front many times. This is what happens. They know.

God, it's not just tech, though. It's almost every industry. They give you access to someone, and then so you'll fall within their thrall, and then they will try to brainwash you.

Microsoft was as good as this as anyone. But you're right, in terms of going back to your original question, I promise your question was- Financial journalists are starting to take, I think, a better look, a deeper look at doing the math. Right?

And what doesn't help during these earnings calls is that the reporter really isn't allowed to ask questions. It's only the financial analysts. And talk about hero worship, they tend to fawn over the executives and tell them what a great quarter they had, regardless of the results.

" And I think once these companies go, OpenAI and Anthropic go public, it's going to really accelerate. But you're already seeing pockets of it. You're seeing it with, extent, Meta will be a great example, I think Microsoft to a lesser extent.

But yeah, there's going to be some pushback. It's like we keep hearing this narrative that this is going to be this absolute boom, and there's no doubt it will be, but we want immediate gratification, immediate results, and this doesn't fit the story construct. Mm-hmm.

So there is going to be some hell to pay among these companies in terms of their investors. There's going to be some brutal results coming up, brutal reaction results. Right.

So Fred, having seen this play before, and when these issues came up back in the days of Bill Gates and crew, that led to the development of Linux, and it was an open source operating system, which was a nice hobby by Linus and crew until vendors and ISVs started pouring resources into this thing. And when you looked at a lot of the people who were working on those projects, they happened to have full-time jobs from ISVs. Well, what was good for the goose might be good for the gander here.

Am I going to see an open source AI agent project where suddenly all the ISVs are donating engineering resources to it because it's going to be guaranteed not to go upstream and steal my business? That's a great question. I'm going to steer away from which places fund those projects to do those things and all of the things that go with that.

But the hobbyist, right? The hobbyist model builder, the ability to use an open source framework. I'm currently working on with OWASP's Agent Control Standard, in which case to help define a way for policy hooks and all of the pieces here that are unrelatable to a specific vendor for that purpose.

Because what do we want to have seen? If folks are going to build things, agents, they're going to train models, they're going to use public weighted models, they're going to do any of these things, they need a ledger, they need truth, they need all these things. Absolutely think that that is the way.

And like you suggested, is there a Linux and a BSD with variants and derivatives and all of that? I sure hope so. And also, you don't need the giant models to do all of these things, right?

These foundation models are absolutely gargantuan. And for some of the tasks that we talk about here, you don't need nearly that amount of firepower, nor do you need to spend that amount of tokens for context, for all the other things that are becoming more insipid as time goes on. So I think there is going to be this boom, and what I would love to see in a perfect world, I'm going to channel my Gina here.

In a perfect world, what I really want to see is everybody has the opportunity to monetize their own models in an ecosystem everybody can contribute with, with some guardrails and consistency and sharing. So, it's not all bad all the time. All right, folks, we're going to move on to the next topic, but watch this space because history has that funny habit of repeating itself.

Anyway, moving on to our next topic, which is also history repeating itself, but there's a report out from Harness talking about how, well, a lot of the spending on AI is wasted, and we saw this in the cloud, right? There was tons of wasted resources in the cloud, and the Harness report suggests that spending waste in AI might be as much as a quarter. The problem is what's spent on AI is a whole lot more than what we spent on a couple of virtual machines sitting in a cloud service somewhere.

So we are talking about hundreds of thousands of dollars, and companies are talking about laying people off so they have enough money to pay for AI. But if we're wasting a quarter of it, Fred, is there something wrong with the math? It's a good question.

There's nothing wrong with the math except the outcomes. " 15 companies pop up to help you understand what your AWS cost looks like. Not necessarily sure how to remedy that because we're not sure the counterbalances.

Now we're asking, like, "Okay, AI, what is this worth? What does this mean? " And it's a hard problem.

If you can't understand not only what's happening with sort of these surreptitious cost spikes, it's happened to us, right? New models come out, we change the flow of the model revenue requirements and so on, and you get caught off guard. It's unpredictable.

Well, the cloud has become predictable. AI's not in this particular case. So if that's also true and we also don't have access to the data, let's imagine that 20% of the time you can identify unexpected costs in a day or two.

Well, that could be the difference between $10,000 and $50,000 in some very, very large organizations. In small organizations, a Four to $10,000. Massive.

And if we think about where the guesswork on the spend, we don't have those well-adapted, well-formed methods of those companies that sprung up for AWS. We have a handful of spreadsheet jockeying, tick token, this thing, that thing, add this up, do the calculus, and then that might match what your provider tells you. That's a hard problem to make a recurring bill spend on something like this, where we're looking at the value of how much money you're spending versus the value you're getting out of that spend.

And if 50% is a good number there, that's the first time in business I've seen that be relevant. So I think my question for you guys is when you look around the neighborhood and you think back to those times and places where those first providers came to bear on what it means, Microsoft, Google, also different models there between these three, and we said, "Oh, you know what? Single cloud platforms, know what?

We're going to diversify. " So three cost models, three different platforms. Mm-hmm.

Five different cost models, 16 different models. Where does it stop? Mm-hmm.

And then I get the pleasure of having to manage all those things, and at some point, the cost of managing all those things exceeds the benefit that I'm getting from the fact that I'm playing a couple of providers against each other. So I might be able to limit that to three, but let's come back to your earlier point. Unlike the cloud, where I might be able to build something that looks like a console to track predictable costs, within all these AI variables, there's nothing that seems to be predictable, so how would I know in any given month what my costs are going to be?

Am I quite literally putting my finger to the wind and hoping, Fred? Is that where we are? I feel like that's true, and my finance guy would probably look at me be like, "Listen, here's the deal.

" And then so, "Oh, hang on, my work is going to stop on the 17th because I've exceeded my quota," right? Or, "I've been rate limited. I've got to ship this thing tomorrow.

You know what? We'll pay burstable fees. " All of that type of behavior is also due to the fact that we're somewhat, and I mean this in the kindest way, held completely hostage in the way that the transparency operates here.

If we've got AI cost policies in place, but I can't enforce them because it literally means I stop doing work, then they're not really policies with effect. It's an accountable problem, and it's an auditable problem, but it is not a spend control. So I think this is the real risk that being caught off guard is one thing, but there's a predictable amount of this kind of a problem and no real way to understand how to remedy it.

Is there any hope to be had from the cloud service providers? And I know on the one hand, they're just hoping people consume as much infrastructure as possible, but even they eventually run out of capacity. And I have seen in the last couple of years that they've been a little more aggressive about teaching people how to optimize consumption of cloud resources because, well, they don't have an unlimited number of data centers, and they're probably falling behind on the new ones that they're planning to build.

So will I see the AWS's and Microsoft's and Google's of the world come around on this equation and say, "Let us help you figure out how to cost optimize the consumption of AI"? Is that possible? I think it's definitely a few years down range for us.

Today it's, "Hey, listen, we have capacity. " And once we break through that barrier, that's a capacity problem. And so when they get to a place where they've built up capacity for all the amazing growth and all the great things that are happening, and I'm going to leave the data centers conversation out of this, when we get to the other side of this, it is, yeah, absolutely.

Now the choice is when there's enough adoption is, "Yeah, you know what? I love Linux, actually. I love small models.

" By the way, I love Bedrock, so no knock against that, but the consistency of that problem is the same old thing we saw with cloud adoption and cloud transformation or whatever we called it. The outcome is the same, which is, I appreciate you curating your spend. Now that that's become a deterministic ability for AWS, you're going to share that with me.

Thank you. But what I'd like to do is I'd like to basically build that infrastructure myself, and then we get into the conversation about, okay, GPUs, compute, are we clouding this thing? What are we doing, right?

Mac Minis don't solve all the world's problems. Hey, Mike, can I invoke... Okay, I think this is the first time we've used this cultural reference, but I'm going to invoke Malcolm X here.

So remember when he talked about the chickens coming home to roost? Mm-hmm. I know this is a stretch, I know, but remember all the talk we heard from the AI leaders about this 30% productivity gain almost overnight?

They didn't say it in those terms, but they led us to believe that this was going to happen almost immediately. And one of the key logic that was left out was that any time you adopt an emerging technology, and we're talking about history repeating itself, there's early stage inefficiency when you're experimenting with it. And with the case of AI, the spending is scaling faster than organizational controls can adapt.

So, I mean, in a sense, it's just an acceleration of what's been happening, as you said, as we mentioned earlier, with cloud and other technologies. So this was bound to happen, and so the survey is accurate, but then again, this is just a A pinpoint in a long timeline, and the numbers are going to change. Am I going to see, Fred, a cottage industry of consultants emerge around this topic who are going to run around and say, "We will help you optimize your AI spend," the same way we saw it in the cloud side?

But is that a viable thing, or are they all just going to be waving shingles at us, but they don't know any more than anybody else? It's a great question. People are already doing this, right?

Because that magnitude is so significant. And I think that's a great thing. The challenge, I think is, to your point, there are some gaps on the predictability part of this that even folks that are spinning up consultancies are going to have a hard time quantifying.

The bigger part of that problem is we don't have aligned incentives. So it's not in anybody's best interest to reduce your token count. I mean, literally anybody.

A consultant that says that, but now the consultant's got a hard job. I have to show you how to get more value with less tokenization, or less spend, or what have you. And that's a hard problem in a place where you're not in an optimization loop yet.

There's not enough maturity here to say, "Well, you know what? I have five or six models that I use regularly. These ones cost more tokens, get you close to the same bang.

These ones cost less tokens. These ones cost less overall," right? And from a provider perspective, are more available.

You're going to get more bang for your buck, but it's going to take you four or five more turns on a conversation to get exactly the answer you want for this particular thing or that outcome here. You got to start looking at all the nuances, and it's a threading a needle problem. And I think that part pushed the consultants hard to shape-shift, but why wouldn't Accenture adopt this?

So in the end of this- Can I- Go ahead, Jim. Oh, Mike, can I ask a quick question? So I think you wrote the story.

So one thing I was trying to grapple with, is this a governance failure or growing pains debate? And I'm wondering- Policy ... policy without enforcement is dead weight, right?

So, like three-fourths of these organizations claim to have AI cost policies, but only half actually enforce them. Was that- Right ... approximately right?

And the half that say that they're enforcing them, I would suggest are probably lying. So here's how this is all going to play out in my mind, right? " And then there's going to be a board meeting somewhere in those organizations, and the knives are going to come out because they're going to be about, well, who gets the best models for their project versus somebody else?

And if there's only so many tokens that can go around, and all the tokens are not of equal value, there's going to be a brawl inside these boardrooms over who gets access to what AI models and who gets funded. Fred, am I crazy, or is that how that's going to play? I completely agree with you.

All right. I completely agree with you. And then there's the forecasting related to employees around solving these problems, too.

And I think there's a lot of big questions to get answered. I'm really a little concerned, but also very curious around how some of the socialization of those problems will begin to happen with real material outcomes for business value at the very large businesses. The DDNs are going to have a dialogue about this, and most of them are going to not necessarily be as informed, right?

The average age of those boards, 68 years old. So, how do you make really good, informed decisions? And that's the part that will get really interesting when you decide that you're going to have a cage match and it turns into a Thunderdome.

There you go. It's going to get curiouser and curiouser. But let's shift gears because there's more weirdness in the world.

So there's three reports out this week all related to security, and the first one is that we are indeed discovering more vulnerabilities in the age of AI, and that's probably a good thing, but that also means that they're getting disclosed. And theoretically, the bad guys are out there writing exploits. Yet there's a second report that suggests that, well, maybe the bad guys haven't quite figured out how to do that just yet because we're not quite seeing the level of exploits of those vulnerabilities, and maybe this is something that's going to happen or not happen.

But there's a third report that I would just correlate to this, and this one came from IBM, and they were saying a quarter of the attacks now seem to be AI-enabled, but they are of the deep fake variety, and basically, they're stealing credentials. And so the question comes back to Fred, let's start here. Maybe it's just too easy to steal their credentials and create a deep fake attack, and the bad guys are going to be, they're watching all this stuff about AI vulnerabilities and all the other stuff that's going on, slightly bemused because they're sitting there going, "Why would I go to the trouble?

" Are we going to find more vulnerabilities? Yeah. For sure.

Absolutely. And that's great. Are we going to find more RCEs and ACEs?

Possibly. I think the challenge for me is it's not just whether or not we find more, or we find more really bad ones, it's the speed, it's the automation, and the chaining. And those things become, if we say you find an extra 10,000 vulnerabilities, right?

" It doesn't matter how many, right? It's the veracity and the velocity that really have an impact here, and also how they're disclosed, what they affect. A long time ago, there was a conversation about 1% of the internet is what you're really after in order, as an exploit surface, for you to build something specifically to do that.

Well, with more agents on the, we'll call it the internet, than humans now- Mm ... the likelihood of that happening has gone up dramatically. The implications of that, I think, are what we're really still contending with.

" And I think the challenge that we have here is there's a little bit of whiplash on, hey, all of these new things are now readily available for everybody, and all agents are able to do that. But the challenge is, historically, we have said that adversaries have more of the timing and the basis for advantage than defenders do. And that's why some of these things hold true.

The flip of that would be, and I sort of disagree a little bit with IBM's statement here. Not in the sense the data's wrong, but when you look at what cyber attacks are AI-enabled, and you look at the veracity and the velocity of what those things have done and the number of compromises, it's still not very big. Not yet.

It will be at some point in time, but the defender dome, as it were, since I said thunderdome earlier, defender dome really is about whether or not the core competency continues to increase to match the agent automation, whether it's triage or it's defense or it's whatever. And I think that is why these outliers skew the perspective of whether or not that's actually true. The notion that somebody uses deep fakes and fuzzing with an LLM, phishing for this, vishing for this, there's a lot of great companies doing and demonstrating that work is very viable.

Some of those exploitation vehicles are very much the same. And so even in the Hugging Face example, where those agents got out and then found RCEs for a particular thing, and then once the agents compromised whatever service was available for them, then doing fan out wide, do very low and slow things, figure out whether or not this particular set of vehicles, the things that humans would not do, make those conversations less about the exploit and more about the exploitation. And I think that's the interesting conversation to have.

The vulnerability is an old trope. So sorry, that's a little bit long-winded, but I think the requirements we have is we've really got to pay diligence to what the data tells us rather than the informal surveys and the vendor illustrations and the agnosticism that this deserves. Well, let's come full circle on that because, John, one of the criticisms that's been leveled at Anthropic is that they've been using this cybersecurity conversation about finding vulnerabilities more as a marketing tool to get- Mm-hmm ...

people to buy more of their AI models than it is some noble effort on their part to do something good, because we're going to make for higher quality software. It seems to me, basically, they've convinced the top 30 vendors to buy their models to go find vulnerabilities because the number and the sizes of the patches that are coming out have certainly increased, and the number of people doing that research has not. So therefore, there's some augmentation at work here.

But is there just a little bit too much of a self-serving cup of tea here from Anthropic in all of this? Yeah. It's a boomerang effect.

You live by the sword, you die by it. In a sense, given what just happened with Anthropic and its disclosure last night, Thursday night, in a sense, it heightens and it spotlights the trust. The security issues are now hitting this company.

It thrived on it, and like any other company, it's suffering from its own marketing because eventually, if you keep pointing fingers at other companies and pointing at how superior you are, you, in a sense, put yourself on an island. And I think in the IBM report, which is interesting, they mentioned the overall breach volume hadn't exploded, but the financial impact of AI-assisted breaches was significantly higher. It was like 6 million per incident compared to 5 million on a global average.

And in a sense, this all feeds into this idea that, again, if you present yourself as a white knight or the do-gooder, you are eventually going to have to face a crisis. And I'm sure, you know what? In defense of Anthropic, I'm sure they will handle this well.

They will move on, but no one's perfect, and I think when you set yourself up on a pedestal, especially out here, people just can't wait to knock you off the pedestal. Yeah. Well, and on that IBM report, they've been doing it for a number of years, and my one criticism of it has always been the same, which is the amount of money you lost due to lost business is an inexact science, right?

There's the actual cost of the breach, and then there's the actual cost of the remediation, which is probably a lot lower than the money ascribed to the lost business, which is, I don't know, Fred, sometimes in the eye of the beholder. Is that a fair statement? 100%.

In today's universe, it is far more expensive for you to cease business operations than it is for you to get breached. Mm-hmm. 100%.

Right. Otherwise known as better to be in business and in the game than not at all, right? Absolutely.

I think, what was it Michael Jordan said? " Yeah. All right.

All right, folks. Well, hey, as we said at the top of the show, there's just a lot of weird things in the world, and you need to think these things through, and they're all going to come home to roost sooner than later. " Gentlemen, thanks for being on the show as always, sharing your insights.

Good to see you as always. Happy Friday. Happy Friday.

And thank you all for watching the latest episode of the Techstrong Gang. Please stay tuned for the rerun of the Techstrong TV lineup for the day. Of course, feel free to watch anything this weekend, and we'll see you all again Monday.