Episode Transcript
Omar Santos [00:00:00]:
You know, it was that moment that I knew that, I mean, it took me about an hour to test it and know that cybersecurity changed forever. And a lot of people ask me, is this FUD, reality? Is this a marketing campaign? Is it token economics?
VO [00:00:14]:
From KBI Media, I'm Karissa Breen, and this is KBKast.
KB [00:00:21]:
My guest today is Omar Santos, distinguished engineer in AI security engineering at Cisco and co-chair of the Coalition for Secure AI, Omar was testing Anthropic's Mythos before it even had a name. We talk about why attackers can now get the same AI capabilities defenders are paying for, how public fix can hand attackers a roadmap, and how long defenders can stay ahead as that gap closes.
VO:
If today's episode lands for you, do 2 things: hit follow and send it to one person in your network who needs to hear it. That's how shows like this grow. Alrighty, let's get into it.
KB [00:01:00]:
Okay, so, Omar, when you first understood what Mythos, or Mythos depending on where you're from, could do, did it feel like— how did it sort of— how did you take it as a security practitioner? Like, obviously there's the general consensus, but I really want to understand your thoughts from a practitioner lens.
Omar Santos [00:01:18]:
Yeah, yeah. As a matter of fact, as I mentioned, you know, I co-lead the Coalition for Secure AI, so I actually started working with Anthropic before it was called Mythos. We all knew that this moment was actually going to come soon, but not this soon. And how I reacted was in the technical term of an AI model that we throw a lot is called scaling laws. That means that if you give enough data, enough compute power, and of course you increase the parameters, these LLMs actually will get better. Now you have a cyber-capable model where it doesn't have a lot of those refusals for creating an exploit, for example. that opens a completely different door. So my initial reaction whenever I started working again, even before the Glasswing, just testing the model or a snapshot of the model is like, well, probably it's gonna be just a little bit of improvement of a new model that they're probably testing that is really good on coding, which is the original intent was a really good coding model.
Omar Santos [00:02:17]:
And of course, the architecture has improved, And now that doesn't have, you know, a lot of the guardrails that, that we had before in models, you know, of course, that's where, where it opened my eyes. You know, it was, it was that moment that I knew that, I mean, it took me about an hour to test it and know that cybersecurity changed forever. And a lot of people ask me, is this FUD reality? Is this a marketing campaign? Is it token economics? And so on. Of course, after the Glasswing, everybody, you know, is doing their own products and so on. So I cannot comment on that, but for sure that moment, and it was around the first or second week of March of this year, right? We, again, we knew that scaling laws are real and these models will become better and everything is in the exponential. It's not about the models, it's the technology around the models, the harnesses, the agent skills, how people are actually finding new use cases. for these models to actually accelerate what we actually do on a daily basis. But in cybersecurity, for sure, that moment was something that increasingly opened my eyes, both in the offensive security side, but I'm also optimistic, right? Some of these capabilities can be used for the defensive side as well.
KB [00:03:28]:
And so you said before, not this soon. Did you envision that this was coming down the pike at some stage, though, in terms of what you think about at the end of the week as a security folk?
Omar Santos [00:03:39]:
I was actually expecting it, to be honest, because these models also need a lot of compute, that probably whenever we started to— all these data centers that we're building starting to come online, that that's where the capabilities will actually go. However, if you look at— if you move backwards like 2 years, we hit a lot of walls whenever it came to scaling laws. One was data, right? All the AI labs ran out of data. Well, we overcome that. by creating synthetic data. But at the same time, you may actually say, well, you know, you're creating synthetic data, but it may be based on probably something that is not that authoritative, not that great. So you're amplifying the problem. Well, there were a lot of techniques to actually, to have and generate amazing datasets that, you know, we didn't encounter before.
Omar Santos [00:04:26]:
The other one was around the time that we did test-time compute. And what that means is that If you think about whenever you talk to ChatGPT in the early days or even Claude, it will just give you an answer and you will be praying that hopefully it will be somewhat of a good answer. Well, now the process of actually before giving you the final answer is going over— people think interact, thinking, the thinking phase is actually something that we call test-time compute. The model is actually not thinking. Or self-reflecting. It looks like it, but in the prediction, it's actually going back and forth and again maintaining that conversational history to then give you a better answer, right? Now you fast forward to today, and I mentioned conversational history, that's a little bit different, but to generalize a little bit, nowadays you actually have agents and models behind those agents that can accomplish a task not in a minute or 2 minutes, but we have agents that we orchestrate that we give the task on a Friday and we come back after a long weekend and that's where we get the results. And in some cases we even have to wait longer. And so we knew that the technology was actually going in that.
Omar Santos [00:05:37]:
To answer your question on how soon, probably I was expecting late next year, right? In the use case of cybersecurity, but I was so wrong, right? Definitely Anthropic did amazing work. Now we're talking about mythos in here. Now it's not about mythos anymore, right? Because these capabilities, you know, GLM-5.2 right now, you know, I'm getting amazing results with open weight models, which we also thought that we're going to be 6 to 9 months behind, you know, the labs. But now that time is also compressing. It's really, really hard to predict or forecast, you know, how this will evolve.
KB [00:06:16]:
What do you think really rattles people the most about Mythos? Do you think it's capability or more so how people are going to respond to it, or do you think it's both?
Omar Santos [00:06:28]:
It is a little bit of both actually, or a lot of both. One is the capabilities of the model or the capabilities, again, Mythos, GPT Cyber, all the labs are actually creating their cyber models right now. So those capabilities from a defender side, It's amazing because now I can instrument agents that can find vulnerabilities that humans haven't found before. And that's a little bit of controversy in the industry because a lot of people say, well, probably we're not looking for 23-year-old vulnerabilities. My reaction to that is, why not? We're still using that code, right? There are very critical components in the supply chain that even in the age of AI, hasn't been tested for security ever. And one of them is because, one, budget. Second, you know, the technology wasn't there, some preparatory protocols and so on. But especially in the open source, if you think about the last week, I was actually presenting at Black Hat in a panel, and we were talking about open source and supply chain.
Omar Santos [00:07:29]:
There's many packages, many open source projects that many organizations are using, that truly hasn't been tested for security ever, right? And I give you another example. Even, you know, years ago, we started actually looking at vulnerabilities in components that you will think that many people will take it seriously. And of course, now I'm not gonna throw under the bus, you know, an open source project, but at that time was the NTP project. NTP is Network Time Protocol. And if you think about it, if I mess up with time, your digital certificates stop working. And that means that your financial system stops working. You know, the core routers on the internet will also, you know, be out of sync. So everything is actually impacted.
Omar Santos [00:08:14]:
And again, this is way, way before Mythos, you know, 8, 10 years ago. Whenever we actually started looking at the open source package, you know, that was in the infancy of these tools that now we have today. And we were able to actually remediate. So now fast forward until today, And now we have these capabilities. Now the ability for us to actually not only find the vulnerabilities, but fix them, and then move away from that whack-a-mole of find and fix only, moving into what are the pieces of code that I have to refactor? Do I have to actually bring security by default within the environment? And then what are the other compensating controls that I can put in place that can help defenders, even recognizing that not everybody's gonna be able to patch, you know, that quickly. What are all the compensating controls that I can put within those technologies that can help alleviate that risk, right? So to answer your question and summarize, you know, that answer, you know, straight away, I am very optimistic because now we have the capabilities of these models not only being able to find vulnerabilities and fix the vulnerabilities at a pace that we have never dreamed of before, but also look at architectural components within code, products, open source packages, that now we can refactor at scale, bring security by default practices at scale in an environment, and also look at compensating controls that we can put within that technology to help, you know, prevent vulnerabilities from being exploited.
KB [00:09:49]:
So, do you think then as well, It's been the catalyst perhaps to your earlier point on people refactoring their code, looking at architectural components. I mean, if this wasn't a thing, people may not have been, they may have been considering it, but maybe not as heavily then, would you say? Because now, from what I'm hearing from what you're saying, Omar, would be people now have to, they can't wait. They have to make these decisions and get on with it versus before, maybe you could sort of dilly-dally a little bit more, but now like the pressure's on, we're seeing people. Yeah. doing things about it in the market. So do you sort of see it like there's a silver lining perhaps to some of the, like, yes, it's not the best, but also it's perhaps somewhat forcing people to take a different position?
Omar Santos [00:10:34]:
It is forcing people to take a different position for sure. I think that this was also a chasm or a catalyst that we needed in the industry to sound a little bit of the urgency of, you know, finding vulnerabilities. If they actually get exploited, they can actually lead to some No bad consequences, even though we actually hear this in the news all the time. But since the technology now is putting like in your face, right? It's hard to avoid. However, you brought an amazing point in that question is that, are we finding vulnerabilities now that probably was already found by a tool 10 years ago, but nobody had the time to actually see if there was a false positive or not? And now AI is actually putting it in a different way. And now because of AI, you're also using it to remediate. And the answer is, you know, yes, in some cases you are finding vulnerabilities now that could have been picked up by, you know, an inferior tool or probably an inferior tool, as she told you again, you know, a few years ago. To give you a perfect example of that is static analysis tools, right? I love them and I, at the same time, you know, don't like them a lot.
Omar Santos [00:11:43]:
I love 'em because it actually can find a lot of vulnerabilities, input validation issues, and so on. Not because of AI, you have to actually throw away all those tools, right? You have to also bring determinism into your environment. However, but now with AI, is not only finding that a piece of code, you know, may be vulnerable to, let's say, a buffer overflow vulnerability or an input validation issue. The way that we instrument AI is that after the, Static analysis, you know, is done. Now we're interacting with the live system. In that case, now we're reducing the false positive from a lot to pretty much nothing. The reason that I said that is because, just to give you some perspective, we at Cisco internally, we don't even send things to engineering to fix unless we have evidence that there's an active exploit, right? Because AI can also hallucinate all day long. And they will give you a lot of false positives if you actually just throw it to a repo.
Omar Santos [00:12:41]:
That's one of the things that we did. A lot of customers were asking us, hey, so you're really finding vulnerabilities? Because I also have access to Mythos and I don't, or I find a lot of false positives. And it was because of a harness that we created internally way before that. And of course it has evolved since the beginning, you know, of the harness, but it was around— November of last year, the offensive security team basically created these modalities of not only testing code for a live system, creating exploits, moving into post-exploitation, you know, techniques, chaining vulnerabilities, and then having agent roles that can do the remediation part, you know, that will suggest a fix for engineering. And we were using other models. So for me, whenever I said that I tested, Sure. I tested a few things that time that Anthropic asked me to test that model, but I gave access to that harness. I gave that harness access to that model.
Omar Santos [00:13:40]:
That's where we accelerated things. That's where we actually truly saw the power of the technology. Now, the other thing that we also are using these models that are non-traditional is for analyzing schematics, for example, in hardware. You know, we're a big hardware manufacturer, and so different techniques that are not just code, but can I detect hardware vulnerabilities or think about how a device can be compromised and how can you bypass some protections, you know, that may have been placed in hardware? And then the other thing that we do in that harness is give access to all the previous findings and documentation of how the product is configured in the real world. Right. And for those of you that are listening that are not familiar with some of the core routers of the internet and things like that, there's like 50 ways that you can configure BGP. And it's not like there are different ways that are bad, is that there are different scenarios. These are very, very complex environments.
Omar Santos [00:14:38]:
That's where we actually are also seeing the power of these AI models, is that now you can actually simulate different scenarios and different deployment models. no pun intended, of those technologies as well.
KB [00:14:51]:
So going back to your point around the inferior tool, if we could address some of those vulnerabilities back then, do you think perhaps we may have had a lot less incident breaches if we had addressed those in the— I'm going to call it the olden days now because we're in the newer days. What are your thoughts then to that? And I ask this question simply because I've sat in companies before where everyone's running an Excel spreadsheet and they've got all these thousands of CVEs and vulnerabilities that have been sitting on there for 10 years, and maybe we may get around to it, but then something else comes in that takes over. So it's like, this could have all accumulated to potentially causing all these issues down the line, but no one would ever know because there's so many of them, it's hard to keep track.
Omar Santos [00:15:34]:
Yeah, yes. The new thing here, and as a matter of fact, let me step back. A lot of people ask me, are you finding new classes of vulnerabilities? And the short answer is no. This is all the same, right? Same input validation issues, same cross-site scripting, SQL injection, cross-site scripting. It's not like it's a new type or class of vulnerabilities. And you have to be humble too, not yet, right? However, what it's finding also, and the motion that we're testing, is to find zero days and unknown CVE. So even if you have, you know, those very comprehensive backlog, you still have it. But now you have more, right? 'Cause now we're finding new things at a completely different scale.
Omar Santos [00:16:16]:
Again, going back to the conversation that I had last week, it was Chris Riverson, he goes by Krub, from the Linux Foundation. We are part of this new movement called ACRIDIS, and it's basically the, you know, helping open source maintainers for incident response, and how are we actually gonna get all these findings from the community And again, it's not about mythos because if you think about now open weight models that I'm using currently can find these zero days. It's not like if they're actually gonna be finding. We're getting a lot of really good results with some open weight models in the community and now their backlog increased significantly, right? So, so you did have some known vulnerabilities in that spreadsheet. But now the challenge is that we're finding new ones, right? Not new classes, but new ones. However, the other thing that I want to highlight is that this whack-a-mole of finding and fixes, we're gonna be there for a long time if we continue to just find and fix and report vulnerabilities. So what we can embrace the technology now is that truly looking at the full architecture, of, let's say, an open source project that has a million lines of code, that is impossible for a human to actually review a million lines of code. Well, now AI models can do it for you really easily, right? So going back to your question related to if I had that knowledge, I don't know, 10 years ago, and I could have remediated, will that avoid a breach? 100%, right? Depending on what the root cause of the issue was.
Omar Santos [00:17:52]:
Now, we've been talking about zero days and so on, but there's also misconfigurations, right? A lot of the things can be avoided. You know, an open S3 bucket with complete access to the internet exposing sensitive information, of course, that's not a vulnerability, it's just that you configure it that way. So now AI can also find that at scale. Another thing that we are seeing in the industry, as I talk to other industry peers, is that some of the traditional technologies for finding those misconfigurations or hidden hardcoded secrets or API secrets in code, it introduced a lot of false positive. So the initial reaction from a software engineer is like, oh, this is a false positive. This is my test code that I have in there to test an environment where I inherited that as an example, but it's not a true secret. What we're seeing now is that you can actually instrument an agent and then validate at scale. Let's say you have Again, a traditional tool giving you 10,000 findings, well, now let me verify those findings and give you the 3 that I can definitely exploit that are not false positive.
Omar Santos [00:18:58]:
'Cause it is impossible for any organization, doesn't matter the size, to reduce that backlog to zero. But you mentioned about known vulnerabilities and also misconfigurations and everything. But now the technology that is advancing, I'm optimistic that we're also gonna be finding new ways of reducing that attack surface, even if you don't patch the vulnerability. Like, what are the compensating controls that I can potentially manipulate the Linux kernel to not allow a privilege escalation vulnerability, for example, right? And we have a technology in the industry called eBPF, which as a matter of fact, the company that Cisco acquired called Isovalent, they created. Those type of technologies now are in the hands of, you know, open source projects, you know, the technique or the, the component is called Tetragon, and you can do a lot of compensating controls that will avoid, you know, an attack. Now, I cannot sugarcoat it, the volume of findings, we're gonna continue to be in that motion of, you know, finding a lot of vulnerabilities as we progress, and it's significant, right? It is significant across the industry. And then the last one is, again, beyond the whack-a-mole and beyond the compensating controls and everything that I said, It's actually even making some decisions. Should you even use this piece of code, right? Can I create it that is better with AI? But is AI introducing new vulnerabilities too? 'Cause it can be a cyclical thing, right? Nobody's coding from scratch anymore.
Omar Santos [00:20:26]:
And if you are, you're probably not gonna exist anytime soon. But the scale of how code is being generated right now and is being put in GitHub repositories, being generated inside of companies is just unprecedented. But if we're now introducing other vulnerabilities, because all AI models will introduce vulnerabilities, we need to actually think about what is the software development lifecycle of AI agents. And as a matter of fact, at Cisco, we started thinking about that early last year. We had an internal project called CodeGuard. And if you think about it, it's the agent skills instead of for— or security skills instead of for a human, in this case for an agent. So if you think about how pretty much all companies and organizations skill their employees, you send somebody to a bootcamp or they pass a certification and they learn all these best practices and the OWASP Top 10 and different secure patterns for coding. And then at 2 o'clock in the morning, whenever you're overworked, then you're probably gonna remember or not remember how to not introduce a SQL injection vulnerability, right? And even if the frameworks and tools get better, but they were not perfect.
Omar Santos [00:21:42]:
Well, now if code is being generated with AI, can we actually, instead of sending that person to a bootcamp, can we guide that AI model or the agent, whether it's Codex, Claude code, Cursor, et cetera, to not introduce the vulnerabilities. We proved that it was amazing what we were actually seeing even in the first week whenever we started integrating this project that I call CodeGuard. Well, we open sourced it and we donated to the Coalition for Secure AI. So now OpenAI, Anthropic, NVIDIA, Meta, they're all contributing into the project. And The motto there is that you can actually use this not only for code generation, but before. So your threat model, your specs, to have security by design in there. The coding agents, of course, generating code. But after the code, what are the recommended security review agents with those skills that then will find the vulnerabilities? Anything that actually will miss, 'cause these are non-deterministic environments.
Omar Santos [00:22:45]:
You hit the before, during, and after of code security. And it's, again, I'm optimistic in that, in that way as well.
KB [00:22:52]:
We'll come back to that in just a moment after a quick word from our sponsor. If you're in legal, risk, or compliance, you know the stakes and the spreadsheet sprawl. Vanta makes life easier by automating key parts of your security frameworks, from evidence collection to audit readiness. ISO 27001, SOC 2, GDPR. It's all in one platform built to reduce manual work without cutting corners. Visit vanta.com/KBKast. That's vanta.com/KBKast to learn more. So do you think that we're kind of solving one problem around the CVEs, reducing that to your point? We're never going to have none, but the list will be a lot shorter from what I'm hearing what you're saying.
KB [00:23:42]:
Then have we sort of created another problem around, well, yeah, everyone's pulling code from GitHub, which may have CVEs, may already exist in there, and then we're using that to layer on other stuff, et cetera. So it's kind of like we've robbed Peter then to pay Paul a little bit. We've just transferred the problem elsewhere. And then if so, is this a worse problem than the original problem, or do you think they sort of balance out to be the same?
Omar Santos [00:24:08]:
It depends on the technology, significantly depends on the technology. And I hate to answer your question with depends, but let me give you some better, better examples. Nothing is immune to vulnerabilities. It doesn't matter if it's code that is generated by the latest and greatest model or, you know, somebody that has been a software engineer for the last 25 years. Nothing is absolutely immune to vulnerabilities or hardware. It's not about software. Hardware is the other perspective. However, I'll give you 2 perfect examples that I'm going to point fingers to only Cisco.
Omar Santos [00:24:42]:
Core router of the internet and an IP phone, right? You will say, oh, you know, if I find a vulnerability that is critical, again, regardless of AI, if I find a vulnerability that's super critical, then take ownership of the device or the system. Which one is worse? The core router of the internet that can bring down half of Europe? Or an IP phone?
KB [00:25:03]:
Definitely one that can bring down half of Europe.
Omar Santos [00:25:06]:
What if the IP phone is sitting on the president's desk?
KB [00:25:08]:
Well, depends which president you're talking about and how big their country is. But then don't you think people could say it's horses for courses though? Like, well, maybe like people don't care about that. It's a president of a country that I'm not in perhaps.
Omar Santos [00:25:21]:
Yeah. The reason that I mentioned that is because it's super hard to say, you know, the technology, what will be the impact? Like I mentioned, NTP, time. A lot of people don't think about, oh, if my clocks are not synchronized, I'm not gonna— I'm gonna be okay. Well, but if the certificates from your bank transfers are not synchronized, or those trading floors that are actually doing, you know, billions of transactions per minute, or patients' records, you know, are not being able to just— a lot of people concentrate again into like, oh my God, post-quantum crypto, which is— I'm not diminishing that. That's amazing. I mean, you have to. That's like step zero. But, you know, some software, to answer now your question, you know, maybe, maybe ignore to say, oh, you know, this is actually not that critical, but people that may be using it for critical functions.
Omar Santos [00:26:08]:
Now, going back to the— is it the same problem or not? I don't, I don't think so. And I have to be humble too, because the technology is changing so fast. Right. So a lot of people have asked me, Is it gonna plateau into 2027? Are we gonna find, let's say if all the models get democratized, all the capabilities get democratized, that now I'm running a Mythos-class vulnerability in my phone and it's so cheap that with 2 cents I can find 1,000 vulnerabilities, right? Are we gonna plateau next year and we're gonna find them all, right? Again, the code generation is the other thing. We are also introducing a lot of code, so it's cyclical, but That's why I want for the listeners to truly, the call to action in here is that look beyond just finding and fixing vulnerabilities, is truly thinking about what will be the security by design and the secure patterns that you can also instrument this technology to avoid saying, well, we didn't plateau in 2027, it's now increasing a lot more, or in 2028. I think that we can actually do, do a little bit more to actually bring down that curve to not be in the exponential just like the models are, right? The other thing is we talked about compensating controls, we talked about code and so on. The incident response capabilities that we need in the immediate future also needs to change, right? So with AI, we need to also embrace it in different ways, recognizing that even if I have a patch, Not everybody's gonna be able to patch at the same pace. Again, the core router of the internet will take significant more than a plugin in a browser, or an IoT device in an oil rig in the middle of the ocean will probably not gonna get a patch in a year.
Omar Santos [00:27:56]:
But you may actually have an insider threat there too that can take advantage of a vulnerability that would, you know, no time. And that one is really, really hard. So in incident response, it goes into 2 facets. One is the typical incident response of an organization, an enterprise, like what we call a CSIRT, Computer Security Incident Response Team. You have a SOC and you're going to alerts and you have typical vulnerability management. But whenever you actually are doing things like threat hunting, that you're doing forensics analysis, you know, AI is also helping to accelerate that. We also have to recognize that the pace of you to actually investigate a potential vulnerability being exploited, it also is gonna increase. So what is gonna be the balance between the technology, token economics, your people, the ability for you to actually have AI generate 50,000 hypotheses for threat hunting, but you can only scale for 2.
Omar Santos [00:28:56]:
So that balance is both a science and an art of itself. And then the second element of the quote unquote incident response is what we were talking about earlier, especially in open source. It worries me a lot. A crucial part of the supply chain is that if I report a vulnerability to an open source maintainer, it doesn't matter, you know, where it is. If you think about the word, the 2 words open source, the source of course is in GitHub or GitLab, is open, everybody can see it. Well, if I report that vulnerability and I can give the fix, well, in that fix there's the diff, and the diff is basically the vulnerable code and the fix. I don't even need an AI model to create an exploit. That's my recipe for taking advantage of that vulnerability.
Omar Santos [00:29:40]:
And in that case, if the new normal is a million vulnerabilities per month in all open source projects, then I can do a lot of different, you know, so if you think about Chris last, You know, last week he was sharing that in the Linux Foundation, they did a study. And again, we don't have visibility into all the open source projects, but it's at least, at least 800 million. So if we hit a trillion, you know, because of AI in the next 10 years, what is the new normal? Is 1,000 vulnerabilities or advisories per week the new normal? Well, it's kind of becoming, to be honest, if you think about it. And how are we gonna patch, how are we gonna provide that patch to the rest of the industry? You know, it is a very complex ecosystem that we have. And because of that, some organizations are also thinking about, do I need to rethink the way that I use open source in an organization? And I'll give you 2 examples. And if you're thinking about that way, you know, bear with me and you will appreciate this. I strongly recommend for you not to replace things like crypto, OpenSSL, and so on. If you do it even with AI, you're probably going to fail.
Omar Santos [00:30:55]:
The reason for that is because even if the technology as of today becomes better, but you have to have really good, strong crypto implementations that are really, really hard to actually get it right. If you try to do your own, of course you're probably going to fail, especially if you don't have a lot of resources. The Linux kernel, Kubernetes, you know, don't reinvent the wheel. Those are really stable, mature open source projects and so on. Now, a library that Omar created last week, and then I put a blog post, a lot of people say, oh, you know, I can do X, Y, and Z with this library that Omar, you know, put in GitHub. And then I fast forward in a year, I moved to another job. I'm not maintaining that anymore. But now you're using it in a web application that is, I don't know, managing healthcare patient records or medical device or anything.
Omar Santos [00:31:49]:
It doesn't matter what the— In that case, does it make sense more to not use that library but give it to an AI agent and say, look at what this library or this piece of software is intended to do, do it better and more secure from the beginning? Probably, especially if, you know, if I move on and I'm not maintaining that project. Another element is end-of-life, end-of-support. Whenever I mention end-of-life, end-of-support, because I've been talking about this for years, probably people are rolling their eyes, oh, this is Cisco again talking about end-of-life and support because they have old devices and so on. Well, it applies to open source too. That library that I said, hey, I created today, but 10 years from now it's not going to be supported. There's no really good way to tell if something that is open source is supported or not, right? There's a lot of techniques.
KB [00:32:43]:
What about like SBOMs and stuff like that?
Omar Santos [00:32:45]:
Yeah, even with SBOMs, as a matter of fact, I actually worked very, very closely with government organizations and Alan Friedman, which, you know, kudos to him. You know, he pushed a lot of the SBOM in a community, but even within SBOMs themselves, In fact, right now the only shameless plug is that the community is actually coming up with a standard called OpenEOX for end-of-life vendor support, but at least to actually classify what is supported from a security perspective and what is supported at all. But yeah, SBOMs don't have it in a structured way. You know, again, OpenEOX is actually working with both of the SBOM communities, you know, out there. But, you know, again, regardless, If it's a commercial product, if it's an open source product and so on, we have to have a standard way that it has to be machine readable, that then with AI agents, et cetera, you can determine, should I use this package? Should I use this product? Should I use— what is the inventory that I have? And that's where also I'm optimistic because it's a common problem that the industry is thinking about solving, both again with the OpenEOX standard, with the technology now with even example agents to look at inventory and to generate and revise SBOMs. In that perspective, we can absolutely take advantage of the technology now to accelerate that risk or to minimize that risk. One thing that you also asked me was around if it was more of the same in code, in the open source packages. And again, I mentioned open source.
Omar Santos [00:34:17]:
I don't wanna throw open source under the bus. I am an advocate of open source. I have contributed, you know, for a lot of my career to open source. I have led a lot of open source projects. But yes, it's actually a lot of the same issues that you see, you know, occurring, you know, over and over and over. It's a lot of the things that we're finding now, you know, is what we probably learning in high school of what not to introduce, you know, in software, right? input validation issues, memory issues, things that, again, we also can use the technology to move to safe implementations. If you think about a lot of the code is C, C++ in IoT, embedded devices, now is the time to use the technology to then migrate to Rust and so on. Now, that is easier said than done in many cases, but for sure, that's a motion that at least at Cisco, we're actually doing right now.
Omar Santos [00:35:15]:
Because the technology is allowing us to do that.
KB [00:35:17]:
Okay. A couple of things I want to ask you, going back to the open source side of things. So are people aware of this or do you still think people are just not sure? But then even if they are kind of aware, well, how do I find out? How deeply embedded is this? Is it from 10 years ago? We're still using some old repos somewhere that, you know, the developer hasn't been maintaining since 10 years ago as well. So how are people sort of starting to combat that then in companies?
Omar Santos [00:35:46]:
Most mature organizations are very aware of the supply chain, you know, issues that, again, including open source and commercial too. But in the community, I think that we have to, we have to be a little bit more vocal and spread the knowledge, right? I'm actually thinking, how can I educate kind of my replacement, right? Those newer organizations and individuals that are now coming into the market that may not be aware of these type of problems. Now is the time to probably address that in there. But, you know, to answer your question about is everybody aware of that? Absolutely not. That's the reason that we're talking about here. We're doing a lot of, you know, outreach in the community. Even the folks that are aware, in some cases it is a very, very expensive Security is expensive, right? And now the cool thing is that the technology, and I'm not talking about tokens, right? Because even with open weight models, you can actually accelerate a lot of the things that you have inside of an organization related to inventory, the way that you can accelerate incident response. Again, it's not about just finding and fixing vulnerabilities, but everything around it.
Omar Santos [00:36:55]:
Again, I'm optimistic on that. But yeah, the awareness is definitely not there. Even within the open source community. Again, I mentioned there are projects that they have people dedicated. There are security people that actually are on staff dedicated to maintain the security of some open source projects. Some other ones are even students that are actually putting GitHub repositories that then they put a YouTube video, they put in a blog on how to actually use that library, and then 5 years later, now you're seeing in companies, right? And they moved on, right? They're not even doing that anymore. As a matter of fact, I have actually talked to a lot of maintainers, you know, open source contributors that in many occasions that say, oh, I didn't even know that my package was actually being used in those scenarios, right? That was not even the internal, the intended use at that time. It was like a proof of concept and so on.
Omar Santos [00:37:49]:
And now they're critical systems.
KB [00:37:52]:
So then my question to you would be, with everything that you're sort of saying, has it really, with the mythos, has it really opened up the playing fields in terms of the whole democratization? Because if I can attack, for example, the same way you can defend, is it just going to go forever like this? Or does it just become a moot point to be like, well, there's no point because we're sort of, we're on the same level now. It doesn't matter. Not, there might. no one potentially could outrank the other sort of sign, for example?
Omar Santos [00:38:23]:
Yeah, actually, that's an amazing question. Let me step back and highlight a few things that I mentioned earlier. Whenever you tested Mythos, you know, the model, was that FUD or reality? You know, did that open your eyes? And of course it's yes. And the first thing on my mind was attackers are gonna have this in 6 months. How in the world we're gonna protect the whole? And you fast forward until today, It has been a little bit less than 6 months. And I'm not saying that it's bad that the open weight models exist, but attackers don't need to pay tokens to Mythos or get into a list to now use the technology that also you can remove super easy the guardrails and then create exploits. So I think that, you know, the call to action in here is that if you're listening, don't think that— because you can get overwhelmed by all the news— don't think that this is only hype. This is not hype.
Omar Santos [00:39:14]:
You have to think about what are the new ways that you have to structure your organization, recognizing that the threat actors have already this technology. It's not like they're gonna have it. They already have this type of technology. The second element is the democratization of the technology is also allowing companies that are a lot smaller than Cisco, at least I'm spoiled, right? I am costing the company a lot, you know, using Mythos and GPT Cyber and so on. significant cost, but we can afford it. A company that may have only one security person or may not even have anybody right now, at least they can use some of this technology to at least help defend their organizations as well. Again, I continue to be optimistic that some of the organizations that can embrace the technology actually will stay ahead. What I'm a little bit pessimistic is And where the time is compressing that if we have, I don't know, let's say 6 months ahead that the threat actors, if by the end of next year is 3 weeks ahead from the threat actors, and then the year after that is a week ahead, right? Like even if you look at some of the websites, there's one called Zero Day Clock, and it has the time that it takes an attacker to exploit a vulnerability.
Omar Santos [00:40:38]:
That's not what I'm talking about. That for sure has been compressing for even before the Mythos and, you know, AI models. That graph is actually going now to minutes. But I'm talking about the capabilities of frontier AI models versus, you know, open weight models that attackers can actually use. That time is also compressing significantly nowadays. Everything is catching up.
KB [00:41:01]:
So then with the time, what do you sort of think moving forward in the next 6 months? What do you think the industry's going to do, or what do you believe we're going to start seeing now out in the market?
Omar Santos [00:41:12]:
Yeah, immediate step is go back to the fundamentals. Not because of AI you have to ignore the fundamentals. If you haven't have a good zero trust implementation in your company, now is the time. There's no excuse. You have the technology to accelerate it and so on. If you haven't actually accelerated your secure development lifecycle, Like, you know, these design patterns, you know, refactoring code, making supply chain security top of mind, inventory, I mean, all those type of fundamentals, now is the time to do it. The next step is how can you build resilient, then in this case, agentic infrastructure for you to do incident response, to do other cybersecurity, even outside of the technical element, you know, If you think about how now most organizations are already operating is the days in the life where you hire a pentester once or twice a year, you put those results in tickets, and you put those in a risk register, those days are gone. So even from a compliance perspective, from a testing perspective, you have to rethink that.
Omar Santos [00:42:23]:
That's new. Now, how can you compress that? It's a continuous assessment in code, it's a continuous assessment in your organization, And then your risk register has to be a living and breathing agent force, if you will, to truly recognize and do simulations of what are the risk, the cyber risk within your organization, right? So even in the non-technical, like the threat hunting incident response, but in the compliance elements, your GRC environment, now is also the time to embrace the technology to rethink and restructure those practices.
KB [00:42:58]:
That was Omar Santos, everybody. What stayed with me is his point that a fix can do as much for an attacker as it does for you. Once a patch goes up in a public repo, anyone can look at the difference and see exactly what was broken. And if we're heading towards 1,000 adversaries a week, that's a lot of open doors. So if you're a leader, listening to this, go and ask your team which open-source components your critical systems rely on and whether anyone is still looking after them.
VO [00:43:32]:
I read every reply. If you've got some thoughts on this one, send me a message on LinkedIn.
KBKast - Cyber For The C-suite.
