Security Confidential S20 E5 Anmol Agarwal

Manoj Tandon: Hello everyone, this is your host, Manoj Tandon. Welcome to another episode of Dark Rhiino Security’s Security Confidential. We have a fantastic guest today, but first, you guys have got to hear me tell you to hit the like and subscribe button. We keep getting information that we have listeners, but you guys are not clicking that button, and it helps us out with the old algorithm. So we would appreciate a little love so we can keep bringing you this great content. But now, without further delay, I would love to introduce our guest today. She has been in this business for a long time. She was a finalist in the 2025 Cybersecurity Woman of the World Awards. She’s been honored by SANS. She has specialization in 5G and 6G networks with a focus on AI and machine learning and API security. She has given talks across five continents, including at places like NASA, AWS, and the United Nations, advising governments, businesses, and nonprofits on cybersecurity, responsible AI, and emerging technologies. We are honored to have her here today. It’s Dr. Anmol Agarwal. Anmol, thank you so much for joining us. It’s a pleasure to have you on the show.

Anmol Agarwal: Thank you so much, Manoj. It’s a pleasure to be here.

Manoj Tandon: So give us your background. You’ve got an extensive set of credentials. Were you always going to be a cybersecurity person when you grew up, or where did this journey begin? Or how did it begin?

Anmol Agarwal: Interesting question. So I did not know I was going to do this my whole career. I think AI security wasn’t even something we talked about before. So I did learn I liked cybersecurity in college. I did the traditional track of studying computer science, and then when it came time to specialize, I was always torn between AI and cybersecurity. I picked cybersecurity, but I always liked machine learning and AI. And so then when it came time to actually do my career, I realized AI security is a thing, and so I wanted to do both. All this AI security information has been talked about for many years. It’s just we’re talking about it more, I would say, in the last few years than we have outside of academia. Inside academia, they were talking about it even 20 or 30 years ago.

Manoj Tandon: Well, let’s do a few things because there’s a lot of questions here that come to mind. Since you know 5G and 6G networks, there’s a lot of conspiracy theories out there about, you know, a lot of radiation, data is going to become easy to retrieve. Give us a little bit of what’s the reality here as a scientist who’s worked on this stuff? What is this?

Anmol Agarwal: Yeah. So I’ve worked on 5G and 6G standards. So I can tell you those conspiracy theories are just conspiracy theories. That’s not real. Actually, every generation, all these scientists and researchers that are working behind the scenes to architect this technology and to deliver this technology, they’re all working to make it better. So 5G is the fifth generation, 6G is sixth generation. Each generation is supposed to be better. 5G is like 4G, but we’re just trying to make it faster, more reliable. And then 6G is basically an advancement of 5G, and we’re adding AI to it. So having AI-native networks to allow you to have something like a live football stadium audience, but it’s virtual reality. So let’s say you don’t want to go to a stadium, but you want to experience the game live, or you want to go to a concert live in a virtual reality form. 6G is trying to do something like that. It’s trying to enable this computing power and this AI-powered network so that you can get advanced experiences like that, or more enhanced sensing capabilities. So when an earthquake happens, we have 6G networks that can do additional sensing. So it’s a good thing, all of this technology advancement. And I would say, you know, if you see it on the internet, don’t always believe it, right? So it’s just a conspiracy thing.

Manoj Tandon: You’re saying everything on the internet is not true, Anmol? That’s a surprising factor.

Anmol Agarwal: Especially with generative AI with all the hallucinations.

Manoj Tandon: So with 6G, when you’re describing virtual reality, is this going to require the special goggles like what Apple has or what Meta has to make it a reality? Or is it a different way? How do people get that experience? I’m just curious.

Anmol Agarwal: I think it really depends on what they want to do. So when I was working on 6G, we were looking at different use cases. Let’s say someone wants to play a video game and they want their video game to be faster, then you would use Oculus VR or whatever smart glasses technology you want to immerse yourself in that augmented reality or virtual reality. But I think it really just depends on the application. And so much of this, you know, we don’t know what it’s going to look like yet. It’s like we didn’t know about ChatGPT until it happened. So with 6G, I think a lot of companies are going to come up with their own ideas.

Manoj Tandon: And with 5G, the microwave radiation which was pointed out by so many people, what’s the real story there? Is it genuinely more microwave radiation, or is it really not any more than what we had with 3G, 4G, or what was there?

Anmol Agarwal: Based on my personal opinion, and I haven’t done research on this, but there’s no reason the radiation would be any different than it was for 4G. It’s just like 4G in terms of that. It’s still, you know, a cell phone tower. I personally do not believe in the conspiracy theory, and I don’t see any reason why it would be different than 4G.

Manoj Tandon: Well, I’m glad you corrected some of us on that. And there was, I think, I don’t know if you’re familiar with it, but Signal Seven used to be a protocol, or is a protocol, that allowed a lot of people to tap into other people’s calls or activate their cameras remotely. It was the backend infrastructure behind a lot of telecom companies. Does 5G and 6G change that? Is that still a protocol in use?

Anmol Agarwal: What I can say is that all of the companies are working to make all of these technologies more secure over time. But the nature of technology is we’re never going to have anything 100%. But every iteration or every technology evolution, we’re trying to make it better and more secure.

Manoj Tandon: Is AI making that task easier or more difficult?

Anmol Agarwal: Great question. I think we are looking at it from both ways. So using AI for security, but also securing AI itself. And I think it’s such a challenge because we see that AI can be secure, right? We use it for data analytics, but then also when you add AI to anything, you need to make sure it’s secure as well. So I cannot give you a detailed specification or anything right now, but what I can say is whenever we think about AI in an architecture, we also have additional specifications and requirements to secure that AI.

Manoj Tandon: Securing AI—define for us what does that actually mean?

Anmol Agarwal: So there are different schools of thought on this, but I think when you talk about securing AI, some of the basic cybersecurity principles still apply. So you could think of AI as just a piece of your system. Software security would still apply in terms of cybersecurity, access control, data privacy, and security—cybersecurity principles like that. But also with AI security, it’s about making sure that whenever the AI is doing something or giving you a prediction, you have guardrails in place. So let’s say your AI predicts something, you don’t want to just believe it right away. You want to have guardrails as well. And again, that depends on your application. Sometimes people use multiple models, so it’s like if in a classroom you have multiple people trying to work together to come up with an answer. So you could have multiple models to do that. Or some people advocate for human in the loop, so having people there to check your AI model and see if it’s doing what it should be doing. But it’s, I think, cybersecurity hygiene, cybersecurity principles, but also checking the output and the results of the AI model itself.

Manoj Tandon: How are you going to validate, like for a complex data analytics problem, unless you have predefined solutions, how do you validate the result? That if you’re putting it on a novel approach with a mountain of data?

Anmol Agarwal: There are different techniques that we use in AI and machine learning, though it really depends on someone’s use case or application. But I’ll just talk about one way, kind of researchy, I guess you could say. So there’s this idea of an ensemble model. People who know machine learning and AI, they know what I’m talking about. But basically ensemble models means we have multiple machine learning models working together to come up with a solution. And when you think about how do you validate it, typically it depends on how you train your AI models. I’m sure everyone has heard now, you know, AI is only as good as the data it’s trained on. So it’s a combination of training it on proper data, but also looking at ensemble models could be an option to see multiple results. And then they could vote or you could decide what result makes sense. And then you also could do something like outlier detection or anomaly detection.

Manoj Tandon: Okay.

Anmol Agarwal: So removing anything that’s different or an outlier. But based on my experience, I think it really depends on the use case. So I can’t give you a magical solution, but this is just some of the things you can do.

Manoj Tandon: You use the term responsible AI. Define that for us a little bit.

Anmol Agarwal: Yeah, so responsible AI—how do we use AI responsibly? Now, people, depending on who you talk to, they will all think this means a different thing because of the word responsible. But in the world we have many different standards now for responsible AI or secure AI. So responsible AI standards could include the NIST AI Risk Management Framework that’s in the U.S. The EU has GDPR, the EU AI Act, and then different countries are adopting their own responsible AI standards. ISO as well. They have ISO 42001. That’s also for responsible AI. So basically, when I think about responsible AI, it’s basically secure AI and ethical AI. Ethical AI means that the AI is fair, which when we think about it scientifically, it means we train the AI on diverse datasets and we ensure that in certain use cases it won’t discriminate against someone.

Manoj Tandon: See, those guardrails are nothing but weighting factors on various neural net participants, if you will, various neurons in that. And if I take that AI and I deploy it locally, I can strip it of those guardrails, potentially, right? Because I’m thinking like a hacker. Because I really don’t care about GDPR. That’s a nice quaint little law. Great. But if my intentions are nefarious, that law doesn’t mean much. I mean, by definition, the criminal mind is not going to give credence to those things. If the other side is working with an unhobbled AI, shall we say, and the good guys are deliberately restricted, that doesn’t sound like on the surface it bodes well for us.

Anmol Agarwal: So for the sake of argument, I’m just going to talk about predictive AI. You can actually have guardrails, if you will, but it’s really secure implementation in the AI itself to prevent a hacker or a bad guy from gaining access to data they shouldn’t gain access to. There is a way you can train your AI model to be robust against certain kinds of attacks, like an evasion attack. So you basically train your AI model to anticipate certain attacks. They call them adversarial examples. And so basically you can train your model to be robust, and then if someone tries to trick it—the classic example is like the hacker and they have a sticker on some image-recognition model to trick the AI model into thinking it’s something else—then you can train the model using robustness training to make it more secure. And then there are other ways that you can add security in your AI model, like data anonymization, homomorphic encryption. So you can do all these different techniques within AI and machine learning to make it more secure. When we have guardrails in place in generative AI, it will say, “I can’t do that,” or, “You’re not allowed to do that.” And so then they’ve used other AI models to give them the right answer or to give them the security answer. From a defender standpoint, that is a tricky challenge to solve because you do want guardrails, but then you also want to allow researchers or people that should have access to these capabilities to have access to them. So I know, like Anthropic, they have this way that you can say, okay, you know, I’m a researcher in this space, I need access to this. I can’t remember what the program is called right now, but they are adding these programs where if you’re a researcher or you’re a cybersecurity researcher, you have access to using these capabilities and to use AI for security research.

Manoj Tandon: See, once you open that channel up, it’s available to everybody. That’s the thing, Anmol. Once you’ve put that in, it’s available to everyone. It’s not just for a select few at that point.

Anmol Agarwal: Right, because then how do you really validate people? Yeah.

Manoj Tandon: Yeah. Not only that, but again, if we go with a local running of these LLMs, it’s not that hard to strip those guardrails off. You can do it in the lab. I mean, it can be done. So it’s interesting. I agree with you. If you’re using Anthropic in the cloud, you can make it difficult. But someone who is, if you’re trying to, let’s say, like the example you used was a really good one. If you’re a malware researcher, I mean, there’s not a lot of secrets in how you build malware. And there’s a lot of books on this topic as well. I mean, this is not like some black arts project thing going on. But if you go directly into Gemini or OpenAI’s model, they will restrict you. They’re not going to give you guidance on how to do certain functions. You can sometimes get around that, sometimes you can’t. But I know the bad guys, if they’re going to go about trying to build a new way to compromise the system and build a new piece of malware, they’re not going to operate under the same boundary conditions that the defenders are. And it’s the same thing. Like look at DeepSeek, the Chinese model, right? It’s a very capable model, but it’s been hobbled because if you go and try and ask it any question about Taiwan, it just blows up. That’s by design. They’ve built that leash into the system. But across the rest of the planet, people are building models that don’t have those leashes. So they are purposefully falling behind a world that’s going to out-progress them. That’s another way to look at it. I don’t know what the answer is, but I think, you know, if we let these models go for objective truth, then let it be what it is, you know? Open it up. What if everybody has access to it? Then is it really a weapon?

Anmol Agarwal: Yeah, I see what you’re saying. I think that’s the nature of the security community, right? Yes, we have Anthropic, Gemini, and Copilot, but we could have other AI models as well. You know, they’re using AI to commit crimes or to enable easier ransomware-as-a-service. But from a defender standpoint, it’s really difficult because you’re basically trying to lower the barrier to entry. You’re trying to say if you give access to everyone, then anyone in the world could easily just do it. Yes, everyone knows how to make malware if they read that book, but then it makes it harder for someone to do that, right? Make it harder for someone to commit a crime.

Manoj Tandon: Harder doesn’t stop. That’s the thing about the resiliency of the human spirit, right? People who have intention, and we can argue whether it’s good or bad. That just depends on what side of the table you’re sitting on, right? If you’re the North Koreans using it to rip off cryptocurrency, well, that’s a way for you to fund your nation-state. So they regard that as a good thing, whereas we may regard that as a very, very bad thing, right? It’s kind of a relative thing there. But if we make it hard for the good guys, the one thing that happens is that you don’t increase the quality, or you make it very difficult to increase the quality of the software that is being developed to be cyber-resilient to begin with, from the ground up, and be resistant to a cadre of attacks that are ubiquitous out there, if you will, in the green field.

Anmol Agarwal: I think there’s a lot of ongoing work about AI security, and a lot of researchers are looking at these different attacks. So depending on what organization you’re affiliated with, you do have access to certain models like Anthropic’s Claude, Mythos, or another AI model, which does have that capability. But the defender standpoint is we want to validate the person who’s requesting access so that not everyone in the world can access it, but just make it so that only those who should have access have access to it. It’s like, yes, access control is not perfect, but we still have access control and we still have passwords. But yes, theoretically anyone in the world, if they really, really, really wanted to try, maybe they could brute-force a password. But we try to make it difficult.

Manoj Tandon: API security. What should be the approach to API security in this modern era now? A lot of those APIs were built before the era of AI, and we saw what Mythos is capable of doing in terms of finding holes in it. So if you’re consulting for someone, what are high-level things to be doing here with API security?

Anmol Agarwal: Yeah, great question. So I think with API security, a lot of it is just access control and continuous monitoring. I’m sure we’ve all seen on GitHub when people leave their API key out there in the open for everyone to see, so then everyone can use that API. So it’s really just cybersecurity principles of access control, making sure only people that should have access to it have access to it. So I think even though the technology is changing, some of the fundamental security concepts don’t change. It’s just a matter of making sure you implement those security principles.

Manoj Tandon: That’s a big ask, Anmol. What you’re talking about is fundamental hygiene and fundamental aspects of a zero-trust framework, right? And there’s a lot of companies that don’t have that implemented well.

Anmol Agarwal: And I think the more awareness we have about these issues, right, that’s when people start acting on it and saying, okay, let’s think about zero trust.

Manoj Tandon: How do you get somebody’s attention with that? Because it’s not like the most energetic topic on the planet, right? I’m asking for my own benefit because when we talk to people, a lot of times their eyes glaze over. Zero trust. How do you—what content, what’s the framing, what’s the approach you take so people can actually get the decision-makers serious about this and it’s not just something for cybersecurity people to geek out on?

Anmol Agarwal: I think with decision-makers, with the people at the C-suite, they’re focused on the bottom line and on money. And so framing technology and cybersecurity in a way that they understand it is important. So talking about if this happens or this consequence happens, then you’ll lose profit or something will happen with the company, and just trying to frame it that way. So communicating in that C-suite-friendly language versus, here’s the technical specifications of how an API works, or this is exactly how an attack would happen. Based on my experience, explaining it in examples and different stories so that they can actually relate to what’s happening versus saying something abstract, you know, like bring it into a story or bring it into an example that they can relate to.

Manoj Tandon: You were a finalist in the Cybersecurity Woman of the World Awards. What are you seeing with women who are aspiring to come into cybersecurity? Are you seeing the space being more accepting, or how is this area?

Anmol Agarwal: I think as time goes on, we’re seeing more diversity in cybersecurity in general. And I think we see that there are more women interested in the space as well. Because I know in the past there were very few women in cybersecurity. Even for me when I was studying computer science, in some of my classes I was the only woman. And for me personally, I didn’t really think much of it. But, you know, sometimes people do want to have role models that they can look up to and say, this person is just like me, so I can see myself in that space too. And I think we’re seeing a lot of different role models now that are doing really interesting things in cybersecurity, AI security, red teaming, threat intelligence. So there are many different women out there that are doing amazing things in this space. And I think that’s also encouraging other women to think of this as an option.

Manoj Tandon: We have had some fabulous women on this show that are doing wonderful things in the field of cybersecurity. But I’m talking in general terms, and because it’s such a growing area, we could always use more people.

Anmol Agarwal: Right. We can always use more people. I think traditionally there are a lot of women who did not see STEM as an option for them or computer science or engineering as an option for them. I think now it’s changing because we have amazing women in this space. And I think the media has a role to play as well. So we see that the media is also having more diverse characters in their movies. You know, they have characters now about women being the hackers or women doing the tech work behind the scenes. So I think that’s one thing, is having more spotlight on the amazing women out there that are doing all the technical computer science, technical cybersecurity, engineering kind of work, because there are a lot of us there.

Manoj Tandon: If somebody wants to switch careers into this field, how would you advise them? Where do they begin?

Anmol Agarwal: Great question. I think it really depends on the person and what skill set they bring. But whenever I’ve mentored someone who is a career changer, I always tell them to think about transferable skills. So, for example, one of my mentees was an English teacher.

Manoj Tandon: Wow.

Anmol Agarwal: She could actually make a really great product manager, program manager, because as an English teacher, you are working with students all day. They have their various demands as students and navigating the different classes and grading all the different assignments. So you’re really good at time management, multitasking, and that can translate very well to project management, program management, and that’s an option in cybersecurity. So it’s just, if you want to do something, you just do it and follow your dreams. But everything takes time. It’s about finding what it is you’re good at, what you enjoy, and then translating that into cybersecurity or into whatever it is you want to pursue.

Manoj Tandon: For companies in 2026, is there any cybersecurity advice that you would be wanting to pass on to them that you see as a consistent trend that happens and they still don’t get it right?

Anmol Agarwal: I think, again, it goes back to just following cybersecurity principles and zero trust, access control, continuous monitoring, logging, just simple cybersecurity mechanisms. Just take a moment—more than a moment. Realistically, it takes months in an organization—but just take time to evaluate your cybersecurity posture and look at, are you following zero-trust principles? I think in 2026, my advice remains the same on that. And I think we’ll see all these news stories about AI, AI security. It’s important to maybe monitor it and see, okay, this is what’s happening, but make sure you don’t compromise on your cybersecurity principles and cybersecurity best practices.

Manoj Tandon: I think that’s very solid advice because when you look at the run-of-the-mill, quote-unquote, hackers who are really looking at financial extraction as the end game, good basic hygiene is still good at deterring that cohort of people.

Anmol Agarwal: Yes. I would say that’s 80, at least 80 percent.

Manoj Tandon: That’s right. And that remaining 20 percent, look, I think if you’re a target of a nation-state, well, they’re going to get in. I mean, that’s—

Anmol Agarwal: We can’t do anything about that.

Manoj Tandon: Yeah, you can’t, you really can’t. I mean, if you’re—you don’t get targeted by a nation-state, I guess. That’s a, you know, behave in ways that keeps you out of that mess. But then there’s also, I think, that word resilience. Perhaps understanding what makes your company or organization resilient in depth can really guide you on the remaining aspects of not just maybe mitigating risk, but accepting a certain amount of risk because there’ll always be a degree of risk just from having a web presence, just from being out there doing business.

Anmol Agarwal: Yes, correct. And so that’s when you have your threat and risk analysis in your different companies, and then you evaluate what risk are you willing to accept, what risk are you transferring, and what risk are you mitigating. But that’s something organizations, I would hope all of them, have that process in place. So that’s just something we would want to have, is a threat and risk analysis.

Manoj Tandon: We had one of the CISOs from Caterpillar on the show many years ago, and he’s also part of SANS, and he developed a threat and safety matrix for the SANS Institute. I thought it was a really brilliantly simple way of cataloging your risks, right? It didn’t require elaborate mathematics and elaborate analytics, but it allowed you to catalog what you thought, not just technically, but from a business-process-wise, what risks are present in the organization. And that often gets overlooked. You look at business email compromise. Is that really a cybersecurity problem? To a degree. I mean, it’s a people problem, it’s a process problem. It’s fooling someone into taking an action based on bad data. They’re really not hacking per se. They’re not compromising a system. They’re compromising a person.

Anmol Agarwal: Yeah, I would agree with that.

Manoj Tandon: You know, and so, but hey, Anmol, thank you so much for being on the show. It was a pleasure having you, and please stay in touch. You know, if you ever have anything you would like to share, you’re always welcome.

Anmol Agarwal: Thank you, Manoj.