March rundown: RSAC warnings and Arm's AGI CPU

0:00This March, speakers at RSAC conference from across the cybersecurity industry warned about the latest threats facing businesses. Some cautioned that just as AI agents are becoming an opportunity for leaders, they're also becoming a potential threat vector. Also, this month, ARM unveiled its first in-house chip, the ARM-AGI CPU. What does it mean? And is this a win for UK Tech? You're listening to the IT Pro podcast. Hi, I'm Jane McAllian. And I'm Rory Buffett. This week, we're welcoming back to Ross Kelly, IT Pro's news and analysis editor to unpack some of the biggest news items from March. Ross, welcome to the show. Hello. So you covered RSAC closely for IT Pro. What have been some of the biggest moments from the conference? Well, it goes without saying that AI was obviously a big talking point at the conference this week.

1:02It seems cybersecurity professionals have their work cut out for them. At the very least, there's an opportunity here for cyber professionals to lead AI adoption globally. That was the key takeaway from the opening keynote by Hugh Thompson, RSAC's executive chairman. There's a lot to be said about the fact that cyber professionals are now doing a lot of firefighting. This has been a common talking point throughout the event. Obviously, massive rapid AI adoption projects, enterprises across a range of industries. But a lot of the time, the speed and the velocity of these projects and the fact that AI tools are filtering their way into multiple business domains. There's obviously a focus here on security, governance, data privacy. But a lot of the time it's, I don't want to say slapped dash, but it's almost an afterthought. And I think that's a byproduct of the hype surrounding the technology.

2:03You know, we're three years into this. I'll cycle at this point. But with the arrival of a genty AI, the games changed even more for a lot of data protection and people in the security side of things. Yeah, you've picked up on something I was about to ask about actually, which is AI agents. You know, this has been the top of conversation since the beginning of the really since sort of the latter half of 2025 that this is where business AI is headed now. And naturally, it's now the focus of a lot of the security concerns. If you like, I edited a couple of your pieces and one of the things that came up was the idea of treating AI agents as if they are human. Could you like expand on that a little bit? Yeah, so on day one in Cisco's chief product officer and president, G2 Patel, describe them as digital co-workers. Now, you know, we've discussed this previously, I believe as well, this idea that you need to treat agents as if they're colleagues, so to speak.

3:04And from a security perspective, that means that you need to apply the same, you know, security standards, the same scrutiny as you would with human workers. You know, identity security practices are very much in focus at the moment. You have a lot of, you know, non-human identities operating within networks within, you know, the back end there. And that's a key focus at the moment. It was something that also came up yesterday as well in one of the final day keynotes. You know, the idea that cybersecurity professionals have often looked at deterministic aspects of, you know, systems, platforms, applications. But then, you know, the probabilistic considerations there around how humans interact with those agents, you know, cover both bases there. You know, they're operating in the back end, but they also have the ability to act autonomously. So that's a secondary consideration there that, you know, raises the stakes because when you have agents with deep access to internal data sets, there's heightened risks there.

4:07There's the potential for manipulation, malicious activity. There's the potential for data leakage there. I think it's just you're adding an additional layer of concern and you're worsening headaches for security professionals. If I wanted to be a bit, well, it is RSAC. So I should really like buy into the whole fud side of things that goes with that conference. You can't even give these entities training. They may be behaving in a way that is similar to how a human would behave, but you can't give something that doesn't have human intellect training, which I guess makes them potentially even more dangerous. Yeah, I think there's an unpredictable aspect there in a talk yesterday. Tenable's Co-Chief Executive, Stephen Vince spoke a lot about accountability. And I think this is where when we're getting into chat around autonomous bots working on behalf of humans or carrying out tasks, there still has to be human accountability there.

5:11You know, how are you deploying them? What frameworks and what guardrails are put in place there to prevent misuse or to prevent lenders? We've already saw last week with Meta around an engineer, you know, essentially taking advice from an agent acting on it and that caused an internal data breach. So there's, you know, how humans are trained to use these tools and not taking them perhaps at face values and other conversation there entirely. Observability, identity management, they've been to big talking points at RSAC. And this is where a gen to AI really needs strong guardrails there. You need to know where they are, what they're looking at, what they're accessing. It is such an interesting topic already that we're seeing these kind of logic errors into the fray with mass adoption of AI agents. So I mean, Ross, you've just said it there with some of the access controls that you really need to strictly enforce with AI agents. I remember a technologist recently talking about an example in which AI agents have access to say payroll information or sensitive employee information.

6:18And if you don't know that they have access to that and you haven't specifically set up a rule that information shouldn't be shared even with another employee. It's not an external breach, but it is something that you don't want to be happening and it's sensitive information being revealed in a way that you really don't want or need for other employees to see. It just seems to be a massive headache for security leaders, particularly as we're seeing. I think it's fair to say millions of agents deployed. I mean, Microsoft and other companies are talking about over a billion to be deployed by 2027, 2028. And I think that's kind of because they come as a package. If you're deploying some particularly through SaaS agents, you're likely to get kind of a full package of different agents that kind of are deployed in your enterprise, which just adds to the sprawl of this, I guess. Yeah, absolutely. And I think Stephen Vince talked about, you know, a responsibility gap here. And I think that's important. What you talked about there in terms of agents having access to information that they shouldn't obviously employees interacting with agents that they shouldn't as well as another big talking point here.

7:29And part of this comes down to the rollout of agents and AI tools in the enterprise. You know, if you're talking about a large enterprise, save five to 10,000 employees, how many separate divisions, you know, departments and teams are all going to be interacting with their own individual agents here. So Stephen Vince talked about the fact that, you know, responsibility here falls to different teams. Sometimes you have the data science team owns and operates the models, the ML ops team owns the production pipeline and it owns deployment. Then you have the product team, which is integrating, you know, these agents, these tools, these applications within other products. And then, you know, as Vince mentioned, cybersecurity teams are at the end of the line. And they are having to react to the shortcomings of integration in different departments. There was a really great example you gave were a financial services on unnamed major financial institution. They rolled out an AI tool internally.

8:37This was quite basic content summarization data analysis and a simple misconfiguration essentially meant that, you know, this individual agent had access to financial models, strategic documents, internal communications. That's not what you wanted that agent to have access to, but it still had it just due to a simple misconfiguration. And the best bit about it is, or the worst bit from their perspective, is that it was an agent built by the security team. So I think that highlights, you know, the risks and the confusion and, you know, how convoluted the rollout of AI and agents and businesses has been for a lot. Another major topic of conversation at RSA is the increasing risks to operational technology, particularly as it pertains to critical national infrastructure.

9:37Ross, I was wondering if you could explain kind of what these threats are and how much of a risk this actually is for businesses. Yes, so this was during a panel discussion on day two, Robert Lee, who is the CEO and founder of Dragos, which is an operational technology cyber security firm mentioned. These are obviously prime targets, operational technology is a prime target, particularly for state backed groups, just because of the level of disruption and chaos that they can cause. A lot of manufacturing firms, critical infrastructure firms in the energy sector, even healthcare. You know, these are all prime targets now, purely in his words, just to take down major portions of a country. You know, I think in the last couple of weeks we've spoken about, obviously, tensions in the Middle East and the war in Iran, we've seen the attack on striker. And, you know, this was not for financial gain, this was just wiper malware, purely to cause chaos there.

10:39And I think this is definitely something that we're going to see, it's something that Lee mentioned, we're going to see more targeting off over the next two, three years. And it ties into global geopolitical events as well. Something else that was announced at RSA was a new tool by Google for monitoring the dark web. What is this Russ and how is it expected to be deployed? So Google had a dark web report tool that users could basically type in their details, email addresses, personal information and whatnot that could then bring up whether or not, you know, there's been leaks on dark websites. If they've been involved in a data breach, if their details are out there, they scrap this tool last year. So they announced plans to scrap it and, you know, supports only just ended within the last six weeks, I believe. I mean, part of this was that Google actively said it wasn't providing bang for buck and this builds on that.

11:39Obviously, this is designed specifically for enterprises here, but a big factor again is just a manual input there. So Google, you know, highlighted the fact that a lot of teams, you know, when they're keeping tabs on potential leaks, potential data exposed out there, they're having to continually update, they're having to continually change keywords. And it's, yeah, it's a lot of manual toil and obviously that's where Gemini can help automates the process. You can build a profile of your business that it will know innately and it can adjust accordingly. As time goes on, it seems like definitely a handy tool for security teams for data protection teams. Testing of this showed that it is capable of analyzing quote millions of daily external events and that's with 98% accuracy. So, you know, this is a really effective tool already, you know, in the hands of security teams, this could be, you know, a key differentiator for enterprises.

12:43A couple of months after cutting that dark web report tool, I remember we reported on it at the time and there was, you know, quite a lot of pushback because it was a really handy tool for, for individuals, for enterprises, this is, you know, probably going to be quite a hit. Away from RSAC and back over this side of the Atlantic, ARM announced the ARM AGI CPU, which is its first in-house chip. What's the story here? Yes, so this is obviously a massive move by ARM here, specifically targeting agente AI. Compute power required for what, you know, is expected to be sprawling a state of interconnected agents essentially. You mentioned, obviously, we're coming back over this side of the Atlantic, we're probably going to have to go, you know, round the earth and highlight the fact that these are still going to be produced in Taiwan. So it's not quite the win for UK manufacturing that you'd expect it to be.

13:44Perhaps not a win for UK manufacturing, but it is a win, I guess, for UK technology in a broader sense. Absolutely, and I think when you look at the fact that Meta has already announced a huge order here, that's a massive seal of approval, a big, big seal of approval. Like I said, you know, this is all around agente AI and the compute power required for it. So when you look at over the last three months with the announcement of OpenClaw, I think it's really interesting here around the fact that you're going to have these massive, massive ecosystems of agents. These agents will be running on CPU cores, and they need capacity, they need memory resources. This directly addresses that. And when you think about the fact that primarily Meta, you know, they've made a big pivot into generative AI three years ago. But now, you know, focusing heavily on agente AI, you look at other big players in the market like OpenAI.

14:48Again, a sharp and focus on agente AI there. So yeah, it's, I think it's a sign of things to come. Rory, obviously, you've got your finger on the pulse here as well. You know, a lot of organizations are building their own in house chips, Amazon, Google. So the stakes are raised here and the race is heating up, I think. I think what's really interesting about arms move here is well, a, this is the first in house chip that arm has produced in 35 years. It's kind of a move away from the IP first approach that arm is taken for all this time. But also it's not moving on into direct competition in the GPU market, which as we know as we've covered is heavily saturated, particularly by the likes of Nvidia and AMD, but also Intel added data center level. This is about CPUs, which Ross, as you said, are super important for orchestrating agents. So in its announcement, arm specifically said this is about orchestrating agents.

15:48It's competing with the likes of Nvidia's Vera CPU, which actually is an arm based CPU. So it's, it's entered into a really interesting market here where what it's looking to do is kind of support whatever existing GPU structure you've got in place, whether it's in video, whether it's AMD, or another company, it's looking to provide racks that can specifically trigger different decisions that agents are making, facilitate communication between agents, and that is a bottleneck. So it's kind of an excellent niche for arm to have to have targeted. It's also clear this has been in the works for some time, ever since soft bank required arm back in 2016, it's poured billions and billions of dollars into internal R&D arm. And the fact that in its opening tranche of customer announcements, you've got arm saying that neta is now arms lead partner and customer. So that's a huge amount of data center deployment that you've got just there, but also that open AI is directly involved in adopting the arm AGI CPU.

16:58We're going to wait to see as and when and how and where specifically these are deployed. There have been some indications I've seen coverage around this that it could have something to do with stargate to UK, which has announced last year among that giant wave of announcements in UK compute capacity, but I haven't seen any hard and fast information on that yet. So, you know, kind of yeah, watch this space. In other case, it's a huge move by arm and it potentially opens it up to a really, really lucrative market just as agents are really hitting that kind of inflection point. And I think that also it's interesting what you talk about this being a CPU rather than a GPU. All the excitement has been in GPUs for a while, but I think that CPUs are going to start to see a bit of resurgence. This is the second CPU related announcement I've seen in the past sort of a couple of weeks to a month or so.

17:59AMD have definitely had something out recently as well that's all kind of like, you know, GPUs are nice, but CPUs are where it's actually because that can affect things like energy consumption as well as you just kind of compute and potentially I'm not a buyer, but it wouldn't surprise me if there's a price implication as well in CPUs versus GPUs. In terms of, you know, competition here as well, I mean, how does this stack up against say AMD because, you know, when you're looking over the numbers here with arm, they're saying it's up to 136 cores per CPU. But then when you look at AMD, you know, they're upcoming epochs and processors, which I think could name Venice, you know, that's up to 256 cores. So as I feel like this is obviously a massive moment for arm, but they're going to have to hit the ground running and there's going to be a lot of work ahead because they face massive competition for me and AMD in this regard.

19:05Yeah, I think it will also depend. I mean, Jane, you said there about energy efficiency that's going to be trade offs between performance and energy efficiency and also latency. What has announced is it's already collaborated with super micro on a 200 kilowatt rack design, which is going to be liquid cooled in which there will be 336 of these agi CPUs, which will mean 45,000 cores in that rack. Arm says that that offers a two times performance uplift in comparison to the latest x86 CPUs. So even against kind of CPUs that it helped to design, it's claiming that this is twice as good. I think we are entering also into an interesting space here with kind of performance per kilowatt as well, where it's becoming less of a question of kind of like what's the latest and greatest option and how much performance can we e count per kilowatt. So when you get these giant performance uplift, it doesn't necessarily mean you're doing twice as much.

20:08It can mean that you're doing the same amount with half of the chips, which is going to be an exciting kind of cost proposition for companies, but also frankly, we're staring down the barrel of massive supply chain problems. So it's become a very good time to do more with less. Yeah. Also, I think as much as I was slightly mocking this at the beginning, the conversation around AI, especially in business, has moved on from just talking about LLMs and generative AI and so on to talking about agents and agents are less sort of their little autonomous things. They're not training a massive large language models. So what you actually need for it to perform well is very different from, you know, at the training stage. We are just moving on the conversation is moving on and we're getting to this point. It's like, OK, well, this is wonderful, but what can actually do for us as a company for me as an employee and so on.

21:13And I wonder if the AI CPU sort of fits into that story as well. I mean, it very much that that seems to be where their marketing is coming from as well. Yeah, I think the focus has changed here. We've gone past the sort of move fast, break things, smash everything up. We need compute power. We need GPUs. You can't get rammed for your desktop computer to to a more focused and targeted use of compute infrastructure. And agents are like you said, obviously driving that. And you know, this is the flavor of all the last 18 months at this point. And if you were to ask me, I think this is we're sticking with us now for the foreseeable. I can't see there being another big paradigm shift in the next two years. I think we are deep in this now and we're going to continue talking about it. And we're going to continue going mad talking about it.

22:16I think from arms perspective, the chief executive, Renee has he's going to have to get a leather jacket because he's going to be out there a lot. Well, unfortunately, that's all we have time for today, but Russ, thank you so much for joining us. Yeah, great to be back. I'm going to my bed after working nights covering on a CC. As always, you can find links to all of the topics we've spoken about today in the show notes and even more on our website at itpro.com. You can also follow us on LinkedIn as well as subscribe to our daily newsletter. Don't forget to subscribe to the IT Pro podcast wherever you find podcasts. And if you're enjoying the show, let us know by leaving a rating or review. We'll be back next week with more from the world of IT. And until then, goodbye. Goodbye.