Anthropic Expands Access to Most Powerful Models as South Korean Banks Are Hacked by AI Agent

0:00This week, hackers reportedly used an AI agent to attack global financial institutions, Anthropic expanded access to powerful cyber AI capabilities for defenders, and the compute market is sending competing signals regarding capacity and productivity. For enterprise leaders, the key question remains the same. How do we turn increasingly more capable AI systems into work we can safely execute, support, and afford? As always, we have Worldwide Technology Chief Technology Officer Mike Taylor with us. To make sense of all of it, it's Saturday, October 10th, and this is the AI Proving Ground podcast from Worldwide Technology.

0:36Let's jump in.

0:45So, Mike, we're going to start internationally here. The Wall Street Journal reported hackers used Chinese-developed agent called Artex AI in attacks on several South Korean banks. This is actually one of the first reported AI-powered intrusions into the global financial system, and it illustrates, yet again, how agents can help attackers more quickly and easily discover vulnerabilities. Given the Globify target, how much urgency do you think this is adding to security teams already looking to automate remediation efforts? Yeah, I think it's one of the more publicly known issues with financial services.

1:22We go back, you know, to the beginning of the latest conflict in Iran. And if you remember, there were very specific targeted attacks against the banks and ATM systems and other things to create that level of chaos. And so I think this is probably more on a national stage. It probably also, you know, exists outside of any direct or acute conflict, you know, that's going on with this, but I think it is a marker that, you know, in our opinion, for the next six to eight months, the attackers have the advantage. And, you know, we talked a little bit about last week, you know, open models and how competition is good in this.

2:03And I think along with this attack, you read in the news that you've got other non-frontier based model companies in China that are coming out with their own models, right, building these out, whether they're e-commerce or gaming platforms, I think was another example that came up. And so, you know, in one sense, that optionality is good for, you know, for the enterprise and for, I think, just industry in general. At the same time, when these models are taken in and obliterated with the guardrails and other things removed from them, you know, they're going to be able to wreak more havoc and they work at machine speed.

2:38The counsel we give to our customers and we're pushing for is, you know, there have been organizations doing really good hygiene over time. Like, you know, they're good at patching. They need to get better, by the way. You know, all these things that we've been good at from a process perspective, patching, tech modernization, you know, end of life, end of service life, just kind of motions that are within that. We need to take those good habits that have been functioning at sort of human speed and put machine speed to them. And I think that's the opportunity, call it, over the next three to six months is how do we start to catch up to these things?

3:15And I'll skip all the cliches of, you know, we got to be right 100 % of the only way we're wrong. But at the end of the day, the two things that are changing with the attack surfaces are the machine speed to detect them. And then secondarily, the anatomy of these typically is that once somebody's in, it takes them a little while to figure out. Think of you're landing somewhere, you're like, what do I have? What do I have access to? Where can I get? Well, they're now getting a view of sort of what they have access to or what they potentially have access to even faster. So this idea of someone coming in, laying and waiting, or slowly sort of looking around, they can do that even at a much slower fashion, much more undetected, but more thoroughly.

3:56Yeah, I mean, well, you talk about the need to get better at patching at machine speed. We recently hosted senior IT leaders from a number of large enterprise organizations here at our global headquarters in St. Louis. And at least one of those leaders described current state patching efforts as a high-speed treadmill, which was, you know, kind of funny, but certainly serious. So if the backlog is growing and growing at a rapid pace, how do we decide what vulnerabilities matter most? How are we prioritizing? What would you tell customers? Yeah, I think this is an opportunity for IT and technology leaders to lead.

4:28You know, too often I can find myself in conversations when it's a technology leader of like, well, we need the business to do this. We got to get the business to the table, which is true. I mean, you know, and that's where I think our CEO led message and that approach is so important. This is business and technology coming together. within that circumstance are four to five, I think, very clear examples where technology, you own it, go do it. Go show your organization what great looks like. And, you know, I do think the biggest, one of the bigger changes we need to contemplate there are traditional organization structures where information security, application development, and IT operations were sort of separate groups and silos.

5:08And our approach has been much smaller teams, much more cross-functional, embedding software development with cyber capabilities, embedding cyber capabilities natively into app development, the patch remediation. And what I've been excited about seeing is the software developers that are sort of bringing their own version of a systems design mindset to the infrastructure teams. Within a few weeks, I'm like, I don't know who the developer or the infrastructure operator was. They're using the same terminology. They picked up the same skill sets, and we're propagating, that mindset around AI data and call it software development to get to some of these automation outcomes.

5:51Because I spent some time in that group, treadmill, exhaustion, no end inside. It was a therapeutic point is maybe the best way to put it for the group that was in the room. But I completely understand what they're saying. Desktops, hopefully in your organizations, I mean, hopefully I say this, you know, you're seeing it in ours. You're getting more and more patch updates. You know, used to be every couple months. Now you're getting them every week. And the same is true for a lot of the backend infrastructure. But lead from the front IT. Come on, technology. Like, you got this. These are things you can chase after and you need to, especially in the areas of cyber.

6:27I like that. A little bit of a rah-rah speech. I mean, one of the interesting things I thought about that South Korean attack was the tool that was reportedly used was an open source agent developed for penetration testing. not for bad purposes, for good purposes, by a cyber engineer known online as Autumn. So clearly, I mean, we're talking about capabilities that were built to help find security weaknesses can also help bad actors exploit those exact weaknesses. I have to imagine that's putting even more pressure on SOC teams to gain access to the latest and greatest tools. And along those lines, this week, Anthropic expanded its cyber verification program, bringing together with it a Project Glasswing to give more verified security teams access to those advanced tools.

7:08Just for those that are unfamiliar with Project Glasswing, it was an effort launched earlier this year by Anthropoc to give early access to Mythos in an attempt to secure the world's most critical software ahead of the broader release. Mike, WWT has been a part of Glasswing now for a couple months, I believe, testing and working with these models and seeing what they're capable of. What have we learned about that program and the difference between finding vulnerabilities and actually making an organization, in this case ours, safer? Yeah, I think there are a couple layers to it. First, the frontier model providers, and we talked about the event last week that was here.

7:42We had almost a month ago, we had some folks from the financial services industry that we brought together who were early adopters and kind of brought into that program. I applaud the model providers who had initially sort of done this in a very covert but purposeful way to try to get ahead of certain things that were coming. And I think some of those patterns still exist today where they're in strong partnership within the private sector and the public sector on how do we help defend some of these assets. The parts that are interesting to me on this as I talk to our customers, boards, and other leaders is the initial introduction of this capability was around scanning code to find novel vulnerabilities or perhaps dependencies across different things that just weren't contemplated in the existing tool sets that were out there.

8:33I think code scanning is a part of what needs to happen, but I think you have to – we should be taking a layered approach to that. Your key network technology partners, the Cisco's of the world, the Palo's, the et cetera, Fortinet F5, all of them, they're going to be doing their own. I mean they're part of that glass wing program. They're going through and they're doing that. And frankly, that's why we're seeing the speed of patching and releasing much faster, number one. Number two, though, is, you know, you as an enterprise, you have these devices you receive and you configure them, you integrate them, you extract and introduce telemetry to them.

9:10And I think the use of AI, whether it's the frontier models or even, you know, you could even look at, you know, Sonnet or Opus, you know, or Opus like GPT-Mini. Pick your model based on your use case, but their ability to go in and scan configurations. Look for dependencies that may exist between devices on your perimeter of your networks first. I think you have to have a layered strategy there. And so, yes, code scanning, especially when you're developing code and you're issuing pull requests and you're promoting, you need to be looking at probably one or two tools to make sure you have all those vectors covered.

9:49And then, by the way, you need some form of continuous scanning there because everything's moving around us, right, number one. What I don't hear enough about in talking to customers is, okay, so how are we thinking about the perimeter? How are we thinking about configuration? How are we thinking about secrets management? Those are critically important in the layer cake approach to kind of, you know, managing your surface area. And so the use case initially of code scanning, to me, these models have evolved based on investment from those providers to be more than code scanning. It's really about vulnerability.

10:25And vulnerability takes the shape of a bunch of different forms in an enterprise or large architecture. Yeah. I mean, you mentioned that layer cake approach and certainly a key ingredient there has to be people and talent. I mean, even beyond remediation, another question confronting enterprise leaders has got to be do the security teams they have even have the skills to turn those capabilities into better production. I read a new report from Pluralsight. It puts cybersecurity among one of the primary areas where skills gaps are derailing technology projects, many of which I'm sure are AI related.

10:58That's got to be no surprise to anybody. But across the technology broadly, nearly seven in 10 organizations, according to that report, are reporting abandoning projects because they lack the skills to deliver them. As AI changes what we're asking people and teams to do, how should leaders at other organizations know when their teams are actually ready to deliver on the ambition that we have with AI? Yeah, I think this is where culture eats strategy for lunch on this. And I say that from a standpoint of you certainly are going to go out and do a highly competitive market to try to hire talent.

11:33You also, though, you have to have an environment that is fostering collaboration and you're going to have to create these skill sets. You're going to have to take an intentional investment, intentional time to go out and take the great people you have and make them better. And I talked earlier about this combination of an app developer, an infrastructure person, a sec ops person. And it's like what's most valuable to organizations is somebody who has kind of the context of each of those areas of domain and expertise because they're going to understand best how to implement some of these capabilities.

12:09And you have to have an organization that one is willing to try those changes out. You have to have a workforce that trusts you to do it. And you're going to have to have a leadership team that is overly communicating on why we're doing it and why we're taking these steps to take this on. And at the end of the day, where I see it work really well is we end up or our customers end up with, you know, they go from having, you know, maybe one or one and a half people that were in that area or that level of expertise. And now they have three or four. And so I still think the market around security and the best people in security is going to be one of the tightest markets that's out there.

12:55The ability to agentify, the ability to learn and train on these, though, to me, it is an opportunity to create more of that within your organization. Yeah. I mean, what's funny or maybe more ironic is even when people have these tools and they can use it more effectively, it doesn't always mean that they're doing it in such a way that creates positive business impact. You talk about tokenomics. We've certainly experienced that here at WWT. Just this week, I was reading a note from Apollo's chief economist, and his name is Torsten Slock. He published a blog that noted meaningful payoff from AI investment has yet to appear in economy-wide productivity data.

13:32Specifically, if you care to know, he was talking about a metric called total factor productivity, or TFP. Now, Slack separated two ways a business can get more done. He was talking about adding more equipment, or in this case, AI, or making better use of its people and the equipment. And the analysis says we can clearly see the spending on AI, but we cannot yet see it in that second kind of improvement across the economy. For good measure, Apollo acknowledged that earlier technologies took years to show up in the broader data set. So that was fair game. And then, you know, take that. But then separately, a recent poll of subscribers to The Information found 35 % of organizations investing in AI are already seeing returns at a multiple of what they spent, with another 17 % reporting positive returns, but maybe less than they hoped.

14:19Those are a little bit competing statistics there. What do these competing reports say about how we should be thinking about economic output and organizational ROI when it comes to AI? Because there's so many different ways to measure here. Yeah, I don't think this is a new challenge. First of all, outside of AI, measuring individual productivity in knowledge working context has been difficult. I mean, as someone managing developer productivity, it's like, oh, well, how did we do? Well, we did it with a bunch of abstract things like stories and story points, which were entirely subjective. You know, I mean, we try to be as consistent about it as possible.

14:59And so I give a little grace to the fact that, yes, this incremental investment, you know, on top of people or along with people is is getting a lot of attention and it should. Most of the companies that we're working with, and I would say we went through this journey early on, it is very difficult to measure individual productivity. Right. I'm going to come back to the individuals here in a second, because I think the culture, the trust, the communication is an important piece of what we have there. The playbook we see, though, is you cluster a group of people that are working on individual productivity solutions.

15:35You work with leadership and you identify patterns within those that you want to move to apply to AI solutions. And when you take a bespoke function and you start stitching it together into a business process, which in most cases has to be a bit reimagined here. I don't think there's enough reimagination being applied to how that's done or even looking at like, why do we even do that? We should should we just stop doing that? What is the value? But it's an opportunity to look at it through that lens. Move these individual productivity pieces into a funnel that says, what is the collection and pattern within this?

16:10And then go and apply it. So in our world, that's RFP assistant. It's proposal studio. It's the workforce planning we do within our supply chain capabilities. It's our services solutioning and how we pull this stuff together to reduce the time it takes to solution and be more accurate with how we bring proposals. Those are end-to-end business functions that you can measure in terms of throughput per hour. Overall metrics of volume against workforce or workforce including AI expense as an example. But you have to move to those applied AI solutions to get there. And I just think for the most part the market isn't there.

16:48holistically. You're seeing pockets. If I come back to the individual workforce piece, the trust, the connection with your employees, we recently launched thresholds within the organization. And we could talk more about those. But in the interest of time, it was a developer, sort of a super user kind of level, and then a general workforce level. For tokens. For tokens, right. Exactly. So across the tools that we're using, thresholds. And one, I think primarily to bring a consciousness to this isn't unlimited. So think about what you're doing. I have been blown away at the amount of productive use cases that have surfaced from places that we did not imagine they would, you know, and so I've submitted a lot of approvals to, you know, move folks from the general workforce into a power user category, just based on how they've embedded this into what they're doing.

17:41And so while not defensible to a CFO or a board necessarily, if I took these tools out of people's hands for what I have assessed to be very productive use cases, we would take a step back, not just from a productivity standpoint, but from a culture standpoint as well. And by the way, I'm sure there's some stuff in there we've got to knock out. We're going to continue to work on that. But for the most part, where I see an opportunity to streamline is you and I are building the same thing. We're spending a lot of tokens doing something very similar that, again, should move through that path from workforce to applied AI.

18:16Yeah, that's a practical way of looking at it. I do want to shift to compute capacity here. The infrastructure market seems to be giving leaders a few different views right now on what's to come. On one hand, we're seeing tons of money put into data centers to meet the AI demand. And we've talked about it on this show about how access to compute will very much be a competitive advantage going forward. But on the other hand, and this is, you know, I was reading recently, former BitMEX CEO Arthur Hayes said at a conference in Singapore this week that the industry is building ahead of demand and that the mismatch could produce excess capacity and much cheaper compute by late 2027 or 2028.

18:54Now, Hayes, we should know it isn't alone in questioning whether AI revenue will catch up to the infrastructure spending. But, you know, that's his forecast, not necessarily a fact, how should leaders be thinking about securing the capacity they need today, but while also preserving flexibility, knowing that the economics could change under their feet tomorrow, next week, or whatever it might be? Yeah. So the first piece I would introduce to this is that it's hard for us to even keep some of these numbers straight. Like, if we took at face value what every NIO, NCP, and cloud provider kind of said they had on their roadmaps to go and build out it it you know we'd be talking about i think the statistic was it was like a 54 increase to the u.s electrical grid right over the next three to four it's not going to happen and so i think there is some some confusion and maybe some gamesmanship in this as well around what resources are available who's subscribed to them or kind of who's going to go in and consume them but if you just take that and kind of set it aside the gating function today remains power.

20:00And it is power in North America. I don't know that we really have an accurate view of what's available and being created in China necessarily, but any geography that we're working with on a Neo sovereign or NCP or cloud provider, the gate is power and access to that power. And so I look at this, those things are not going, those are not light speed sort of events. So we have a fast follower circumstance here where I think we're going to be able to watch this play out. And I pay less attention to the accumulation of supply and demand numbers that people are forecasting or projecting out. I pay more attention to what is the current state of the market?

20:44Where are these demands? And we're not working with a single GPU provider out there who isn't subscribed to 95 % plus of the capacity they had. I'll accept the argument that OK, some of this is on for a lot of this is on forward facing demand from OpenAI and Thratpik and others that are out there. We haven't even, though, introduced real enterprise demand to these. And the irony in this a bit is that when I'm working with our enterprise customers who want 50, 150, maybe 500 on the high end megawatts of capacity, we can't find it. I mean, you know, and so what we then work through is a circumstance of, okay, if you have real demand and use cases, what is the circumstance for a private build?

21:34What capacity do you have to do that? Whether it's in your data center or you're partnering with the data center provider. And then where are the opportunities or holes to fill in this massive amount of capacity being consumed by cloud and other frontier providers and figuring out where that fits in? And all to say, I think the answer in this is my answer back to the first piece is I don't think we've seen the enterprise demand really come forward here. It's too early. I think the evolution and maturity, the gap between the open source and frontier models is beginning to collapse on itself. It'll never be zero.

22:07We talked about that before, but the margins are going to get closer. And as they do, I think some of this will require you owning your own destiny with infrastructure on your own premise or within your own control. Yeah. Well, I mean, as enterprise enter the fray here, you know, the capacity may exist, but it's going to be, you know, potentially reserved by others. You know, there was a report out there that I was reading that, you know, to help mitigate that type of scenario, Axios was reporting that former execs from Andreessen Horowitz, Google, Apple, NVIDIA, I think was in there, or at least former execs from NVIDIA, launching a new company called National Compute to pool AI computing capacity from labs, cloud providers, or different chip systems, whatever it might be.

22:48and then allocated through a shared scheduler. The project was a way to share computing for both commercial research and public sector use. I think the business model there is still being worked out, but could pooled capacity become a credible alternative to dedicated reservations and help alleviate some of those issues? Like you say, when enterprises do jump in full force and the capacity is real. Yeah, it's an interesting idea and I think it needs to be available. I mean, I think if you're running an enterprise, you may have some percentage of your consumption where that mode is tolerable and maybe it's more price efficient to be able to do, but you're, you're going to need the always on, always mine, always under my control to the, you know, borrowed and, and, and least, you know, capacity of this.

23:34And so I, I think it's an interesting idea. There's a lot that has to come together there in terms of the software stack. And I don't mean the piece that kind of sits on top of it, but makes it, you know, makes it work. I mean, the ability to, at really low latencies, potentially connect capacity across physical sites, I think we're a ways away from that being a reality. But are there smaller segments of larger distribution, gigawatt level facilities where you can carve off 50 and use them overnight? Yeah, maybe. You know, but on the enterprise use case piece, though, one of the points I wanted to share in this was, you know we're talking primarily about like inter i'm sorry intra enterprise use like us proposal studio everything else like one of the one of the trends that's going to drive this this up beyond cyber and what's happening is we still haven't seen a mature base of enterprises interacting with enterprises using ai you know and sort of this idea that you know today we have service-oriented architectures and technology that integrate orders, they integrate dates, they, you know, they share data back and forth.

24:45You know, you think about credit card processing, you think about, you know, as consumers, we do online shopping, other things. Getting to a point where these agents are understood, secure, certified, almost like, you know, like a DNS registry for an agent to interact with these different, that's a whole another level of usage. And we didn't even talk about robotics and all the other things that sit on top of it. So I had to get those things out of my head just in terms of, it isn't just the internal use cases these enterprises are going to drive. It's the stuff that's on the six and 12 month, I think timeline horizons too.

25:19Yeah, no, just scratching the surface for sure. Well, Mike, appreciate the time as always. Always love having you here in studio. We'll talk to you again next week. Yeah, sounds good. Thank you. All right.

25:32.