Authorized Intelligence cover art

Red teaming with frontier models

November 5, 2026
Ep 1

This week on Authorized Intelligence, Cal Al-Dhubaib speaks with Nidhi Aggarwal, Former Chief Product Officer at HackerOne⁠, about how frontier AI is transforming cybersecurity red teaming and vulnerability discovery.

23 min
Red teaming with frontier models with Nidhi Aggarwal

This week on Authorized Intelligence, Cal Al-Dhubaib speaks with Nidhi Aggarwal, Former Chief Product Officer at HackerOne⁠, about how frontier AI is transforming cybersecurity red teaming and vulnerability discovery. They discuss how AI is increasing the scale and sophistication of security testing, why validation and remediation are becoming critical bottlenecks, and why organizations need continuous security practices that combine AI capabilities with human judgment.

What You’ll Learn

  • AI is expanding vulnerability discovery. Frontier models allow researchers to examine larger attack surfaces and entire code bases, uncovering vulnerabilities that previously exceeded human capacity.
  • Validation determines real exposure. A vulnerability must be tested against production environments and existing security controls. Aggarwal explains that thousands of discoveries may represent only a handful of truly exploitable risks.
  • Continuous security is essential to cyber resilience. Organizations need an always-on cycle of discovery, validation, prioritization, remediation, and retesting rather than relying on annual penetration tests or one-time assessments.
  • AI increases the need for human judgment. Security teams will increasingly review and govern AI-driven work, requiring stronger cross-functional collaboration and an understanding of business objectives.

Key Moments

  • 00:00 How is AI changing red teaming?
  • 03:00 Why does vulnerability validation matter?
  • 04:00 What makes an organization cyber resilient?
  • 08:00 How fast is AI changing security?
  • 10:00 Where are organizations falling behind?
  • 17:00 What should organizations prioritize now?

About Nidhi Aggarwal

Nidhi Aggarwal is the Chief Product Officer at HackerOne, where she leads the company’s product strategy and platform vision. With a background in technology, product leadership, and entrepreneurship, she brings a perspective on how AI can scale security capabilities while ensuring human expertise, validation, and governance remain central to effective cybersecurity.

Listen & Subscribe

If you enjoyed this episode of Authorized Intelligence, subscribe wherever you listen to podcasts and leave us a rating or review. Have a question or a topic you’d like Cal to explore? Get in touch with us through  authorizedintelligence.fm.

Cal Al-Dhubaib: So my, uh, question to you, um, at HackerOne, you’re really focused on the evolution of red teaming. Mm-hmm. What’s changed over the past few years, especially with the emergence of, of frontier AI capabilities?

Nidhi Aggarwal: AI capabilities? Sure. Uh- There are two major things I’ll point out. I mean, first of all, uh, security and, uh, red teaming is constantly changing. Uh, but with the emergence of frontier models, the biggest thing that has changed is that security researchers are at the forefront of adopting models.

Cal Al-Dhubaib: Scale.

Earlier, it used to be about, uh, eh,

Nidhi Aggarwal: the second thing that has changed is the scale.

Earlier it used to be about, uh, right, okay, can I get in? What can I get access to? Now it is about the emergent behavior of the entire system. What can the model do? What can the model ac- Uh, how is the entire system behaving in a non-deterministic way? And that is changing. This

Um, [00:01:00] I think it will get worse before it gets better.

Cal Al-Dhubaib: before

Nidhi Aggarwal: And, and, uh, I’ll explain why.

Uh, we are going to discover more vulnerabilities than ever before, because earlier it just wasn’t possible for us to go look at the entire attack surface, go look at the entire code base, because it was human limits. Now we can go, uh, and do that. AI is part of the solution. Yes, AI is, uh, right, uh, both terrifying in terms of creating this problem, but AI is part of the solution because it provides us the s- speed and the scale to be able to find the vulnerabilities faster than they can be exploited- Got it

and fixing them.

Cal Al-Dhubaib: Have to do though is say which vulnerabilities matter, what is the real risk- Sure ... what is exploitable, so that we are prioritizing our resources to the vulnerabilities that matter, and doing this continuously. We can’t make this a one-time exercise, and AI helps with that. Absolutely. And, and we see this as well, this shift towards continuous security, continuous governance, and it, it’s not enough to have 90%.

It truly has to be always on. It

Nidhi Aggarwal: has to be always on

Not only are the models able to find more, [00:02:00] because there’s a lot of emphasis on, “Hey, the models are finding more.” Sure. They’re able to find more sophisticated vulnerabilities,

Uh, so that makes it different. Uh, the other part is, yes, when we are finding more, the bottleneck shifts to what is the real exposure and how should we fix it?

Got it. We ourselves weren’t able to keep up with this, and this is where AI comes into the picture,

Cal Al-Dhubaib: The, the cost of not using AI is much greater. It

Nidhi Aggarwal: is much greater, because now you know about it, and now you have to fix it. Um, the third part is we do have to evolve. Model is only a part of security.

You have to evolve your harnesses, the governance, and how are you going to use the model? What are the safeguards? What are the ways in which you’re going to optimize the performance? Because models still produce a lot of false positives. Sure. And w- like, there’s a lot of talk about A- AI safety. So that harness, uh, around it becomes really important.

Got it.

Cal Al-Dhubaib: And so, so it sounds like these models are changing [00:03:00] the sophistication at which we can uncover these vulnerabilities- Mm-hmm ... especially lower level issues that can be chained together. Um, and if you’re going from uncovering, let’s say, a hundred bugs to now a thousand bugs a run, even if you’re ninety percent accurate, that means you’re, you’re getting about, you know, a, a hundred vulnerabilities that you’re trying to investigate that are just noise.

Yeah. So, so how do you, how do you deal with that prioritization? So-

Nidhi Aggarwal: So, uh, what you have to do is the validation layer becomes really important, and what I mean by validation is you have to test that... It is a, uh, right, a vulnerability is theoretical unless you truly test it, right, against your production environment and say, “When I have all of my security controls on, is this truly exploitable?”

So you have to test for exploitability. Uh, is this code truly reachable? Is this path truly exploitable? Is, is this going to be stopped by a WAF rule? ‘Cause that’s how you [00:04:00] prioritize. And out of the thousands, you might even find that there are only 10- Sure ... that are truly exploitable, and th- that represents true exposure, and those are the ones you prioritize

Cal Al-Dhubaib: th- that represents true exposure, and those are the ones you prioritize for remediation. Got it. So the, uh, burden of proof you have to ask is discovery alone is not sufficient. Discovery has to be combined with validation so

Nidhi Aggarwal: the, uh, burden of proof you have to ask is discovery alone is not sufficient. Discovery has to be combined with validation so that remediation can keep up. Yeah. And it has to be that continuous loop. So

Cal Al-Dhubaib: in this world, it’s no longer if. It’s certainly no longer if. No. But when How do you interpret, um, the meaning of being a cyber resilient organization in this new reality?

Nidhi Aggarwal: Uh, I think you started the question with, uh, it, uh, it’s if and, uh... It’s not no longer if, it’s when. Um, I’d say cyber resiliency,

Cal Al-Dhubaib: no longer if, it’s when.

Um, I’d say cyber resiliency, uh,

Nidhi Aggarwal: a few [00:05:00] things. One, uh, some things don’t change. Be brilliant at the basics. We have to do the things that we’ve always known because AI amplifies the things that, uh, we were able to live with or sort of get away with.

Uh, and now we just can’t. Uh, so for example, right, if you, uh, were not really diligent about your access controls or, right, building, uh, isolation, network isolation, uh, right, so, or zero trust, uh, architectures, you’re really exposed. So, uh, be brilliant at the basics. Uh, two, uh, invest in this continuous security loop.

Uh, constant, uh, discovery, validation, prioritization, remediation, retesting, and where... And this loop has to be always on- Yeah ... because that’s what makes you resilient. Because you, you are going to be developing and deploying more and more, uh, code, so your attack surface is gonna grow. So this [00:06:00] continuous loop has to be, uh, part of it.

And there is no one silver bullet. While you use AI for the scale and speed, human judgment and creativity remains an very important part, uh, to, one, provide the... Okay, what... Where does AI, uh, push the limits, but where do the humans take it? Sure. I started with the OpenAI, um, example. There, model took it only so far.

It’s the humans, uh, who were the creative ones. Uh, but also for the safeguards. I call it humans over the loop.

Cal Al-Dhubaib: humans over the loop. So,

Nidhi Aggarwal: you have to design for what are the safeguards, how you de- you deploy it, what’s the governance. So you have to, uh, for a cyber resilient org, it’s not just about the model. It’s everything around the model.

It’s the entire system, prompts, data, governance. You have to invest in those.

Cal Al-Dhubaib: have to invest in those. I, I love that perspective so much because there’s almost this myth- Yes ... of we’re gonna use AI to increase our productivity everywhere, we’re gonna [00:07:00] use it to improve our security, and that somehow translates into fewer jobs to be done.

And, and I think what we’re both seeing is it’s, it’s growing the amount of work to a point where you can’t hope to do it without AI, but you also certainly need all the humans and maybe some more.

Nidhi Aggarwal: You, you will. I think certainly Jevons paradox will play, play out here. Yeah. Uh, that you are going... It’s going to be, uh, easier to find one vulnerability, but we are now going to try and find more vulnerabilities against an ever-expanding attack surface, and this example will apply to every aspect of security,

Cal Al-Dhubaib: the- there’s so much that I’m constantly changing my mind about as I’m watching this AI story unfold. And so for our listeners, um, especially given your vantage point, you, you were so close to the frontier labs, what have you personally changed your mind on within the last 12 months when it comes to, uh, frontier AI and security?[00:08:00]

Nidhi Aggarwal: I think I’ve personally changed my mind around the, uh, pace of change because the pace of change I anticipated, like the capabilities that we see today, I probably would have expected them, right, slightly later. Yeah. And what does that... that has done is the, um, I think for me, uh, what is that has changed is the need for different communities to come together- Yeah

is pressing now, uh, in terms of like this, uh, for example, there’s a lot of conversation around AI safety. Sure. And I think that conversation is happening in, uh, right, uh, little islands. Uh, there is the AI, uh, development community that is, uh, talking about AI safety separately from the security community and from, say, the, uh, legal and regulatory community or from the, um, the philosophy even or morality and ethics community.

Yeah. But bringing that [00:09:00] conversation together needs to happen now at a faster pace because I’m an eternal optimist. I’m a tech optimist, and I think like AI is going to really help us be, um, right, have a, build a better future for all of us. But it would not happen by happenstance. There is work that needs to get done, and it is upon all of us to, uh, to come together and build for that world, and I had anticipated that the pace of change might be slightly slower, so we might have more time to adapt, uh, to that, uh, change.

Um, but, uh, I think we need to adapt to it faster and bring, come together faster.

Cal Al-Dhubaib: that, uh, change.

Uh, but, uh, I think we need to adapt to it faster and bring, come together faster. That’s a fascinating perspective, and it reminds me actually of a quote from one of the sessions I, I was at yesterday, uh, from, from DeepMind. And, uh, the gentleman said was, uh, “This is not a forecasting problem. This is a design problem.”

Right? It’s no longer guessing- It’s no longer guessing ... when is it gonna

Nidhi Aggarwal: This is no longer guessing. We’re actually doing it. Right. When it’s, it’s here, and it’s now. And

Cal Al-Dhubaib: and now we have some choices

Nidhi Aggarwal: to make. Exactly.

Cal Al-Dhubaib: I love this. So [00:10:00] where, where are organizations today the least prepared for this new

Nidhi Aggarwal: reality? I think two levels. Uh, one is clearly operational, um, because security is, uh, for most organizations, their security infrastructure is still designed for a world where security was designed for human speed, uh, human scale, and also a lot of the processes are for a one-time event, uh, like a compliance event.

You do a pen test once a year, uh, right? It’s not designed or retooled for this continuous security-

Cal Al-Dhubaib: Absolutely ... um, right, operational layer.

Nidhi Aggarwal: Absolutely ... um, uh, right, operational layer. Uh, so that’s where they’re, uh, least prepared. Uh, the other one is, uh, this, uh, um, mindset of, uh, if I, uh, right, uh, can get some, uh, level of trust about, uh, uh, the model, I’m safe. Yeah. I think this, uh, concept that we talked about, model alone- Yeah ... isn’t the entire attack surface. You have to really think about everything [00:11:00] around the model and the system, and the system isn’t going to behave deterministically. So that emergent behavior where, how will the system behave when you give it a different prompt or different set of, uh, right, environmental variables, that is something that organizations are not prepared for.

Cal Al-Dhubaib: organizations are not prepared for. And something that you had mentioned to me earlier that I think would be very fascinating to explore is we, we trust the frontier labs to be great at frontier models.

What are your thoughts about their security capabilities?

Nidhi Aggarwal: I think they’re trying, but that is not what their expertise. And so security companies or security community has, uh, right,

Cal Al-Dhubaib: the

Nidhi Aggarwal: domain expertise, uh, to be able to...

that needs to come together, uh, with the frontier labs. So I think we... it’s the, uh,

Cal Al-Dhubaib: sometimes

Nidhi Aggarwal: I sometimes call it the halo effect.

Cal Al-Dhubaib: Yeah. Right? Uh, where,

Nidhi Aggarwal: where, uh, if, uh, something is, [00:12:00] if, uh... And that’s a, right, normal, uh, tendency. If, uh, certain, uh, if you have certain capability, uh, that you’re really great at, it’s assumed that you’re great at the second, uh, capability as well.

Yeah. That might not be as true, and, uh, you might, uh, benefit from, uh, domain expertise, uh, in that, uh, second capability.

Cal Al-Dhubaib: Fair enough. Fair enough. And it, and I, I think at the very least, there’s a lot of assumption that these, these models come with guardrails, and, and that’s necessary, but not sufficient, um, and doesn’t come with the breadth of context needed to truly operate securely within the confines of our own organizations.

Nidhi Aggarwal: That’s absolutely true, and that is one of the things we, uh, are doing as part of these partnerships. Because when we get access to these models, uh, especially for defense, uh, they n- don’t have guardrails so that we can actually, uh, do the [00:13:00] defense work, and models don’t, uh, right, uh, have refusals when we are asking them to do cybersecurity work.

But then, uh, our investment is in building all of these safeguards and guardrails and harnesses that, uh, keep the model safe so that when they’re deployed for defense, they do what we intend for them to do and not what we don’t intend them, uh, to do. So that is the work that needs to be done, and this is where the security expertise comes in.

Cal Al-Dhubaib: true, and that is one of the things we, uh, are doing as part of these partnerships because when we get access to these models, uh, and especially for defense- Yeah ... uh, they don’t have guardrails so that we can actually, uh, do the defense work, and models don’t, uh, right, uh, have refuses when we are asking them to do cybersecurity work.

But then, uh, our investment is in building all of these safeguards and guardrails and harnesses that, uh, keep the model safe so that when they’re deployed for defense, they do what we intend for them to do and not what we don’t intend them, uh, to do. So that is the work that needs to be done, and this is where the security expertise comes in.

So we talked a little bit earlier about the need for more human work. Mm-hmm. Two years from now, what does the workload look like for the average security team? So I think, uh, the

Nidhi Aggarwal: I think, uh, the av- uh, the workload for the average security team in this continuous, uh, security era ... right,

Cal Al-Dhubaib: uh,

Nidhi Aggarwal: uh, shifts from doing the work to reviewing and governing the work that is being done.

So, uh, reviewing, um, like [00:14:00] what are the kind of issues that are surfacing and, uh, what needs to be done about them, what are the governance issues around those. So that’s the shift in mental model that will occur for the security teams.

Cal Al-Dhubaib: in mental model that will occur for the security teams. And do you see this requiring new expertise, new roles, uh, closer collaboration with different groups? Like, what changes?

Nidhi Aggarwal: So, uh, I do see, uh, right, a higher level of human judgment. The expertise required, uh, would be like, like if I’m going to review this, I need to have that judgment, and I need to have that expertise. Uh, so I do, uh, see the need for more critical, uh, judgment, um, in the security roles. Um, in terms of governance, this would be cross-functional because now I have to understand what is happening in the rest of the world, what should be...

Because security’s role, right, I always said security’s role should not be, uh, to say no. Yeah. It should be how can you safely say yes?

Cal Al-Dhubaib: So to

Nidhi Aggarwal: So to [00:15:00] be able to safely say yes, you need to understand more of what is the other part of the org trying to do. So there’ll be more cross-functional collaboration and more need to understand the other parts of the business.

This,

Cal Al-Dhubaib: uh, this kind of goes back to the joke that the most secure way to use a laptop is to disconnect it from the internet-

and turn it off. Yes.

Nidhi Aggarwal: and

Cal Al-Dhubaib: And so if you’re gonna be governing the use of these, these tools, you might actually also need to understand how they’re being used in day-to-day business, so that as you’re, you’re making these judgment calls of what gets escalated, what features get turned off- Yeah ... um, you’re, you’re not actually disrupting the business as a result of these decisions.

Nidhi Aggarwal: And you’re allowing for experimentation because one of the greatest benefits of AI is that it speeds up the pace of experimentation. And because you can experiment more, uh, you will, uh, get to a better idea.

Cal Al-Dhubaib: the pace of experimentation. And because you can experiment more, uh, you will, uh, get to a better idea. So there’s, there’s two forces here. You’ve got your, your business and your tech [00:16:00] leaders on one hand, and then you’ve got your security leaders on the other hand. Uh, for our listeners in the audience on, on both sides of this, um, what is one question that they can bring to the other side, uh, to help them advance this conversation?

Nidhi Aggarwal: conversation? Um, for business leaders, it, it would be, um, “I wanna do this, and help me, uh, figure out how can I do this safely.”

Uh, so, uh, don’t just go ahead and do it because there are real, uh, security risks. And for security, uh, right, I think this is the... We’ve already talked about it a little bit. Yeah. It’s, uh, uh, what is the outcome you’re trying to achieve? Don’t start with the process, uh, that, uh, here are the steps to the process, and we need to have this gate and this gate and this gate. Yeah. Be a partner and ask for what outcome are you trying to achieve, and I think that will lead to a better conversation on what are the safeguards that we can have [00:17:00] that lead you to that outcome.

Cal Al-Dhubaib: outcome. I love that. So, so essentially to your business leaders, treat your security team like an in-house consultant. Yes. This is not the department of no anymore. No. It is not. This is, this is your best partner- Yes ... for figuring out how to move forward, and then-

Nidhi Aggarwal: S- And safely

Cal Al-Dhubaib: And safely ... and then you’re telling security- Yeah

act like a consulting partner, right? Try to understand the outcome. Try to be outcome-based. Brilliant. Brilliant. So if you could get every company to do just one thing different

Nidhi Aggarwal: right now- Mm-hmm, mm-hmm

Cal Al-Dhubaib: to help improve our overall security state, what would that be?

Nidhi Aggarwal: be? I would say that stop treating security as a one-time assessment, especially AI security. Start investing in building this continuous security loop because, like you said, it’s a question of, uh, when, not if, you’re going to have to build this resiliency towards this always have to be defending against these kind [00:18:00] of, uh, uh, attack vectors.

So start investing in this continuous loop. Start investing in a continuous discovery loop. Start scanning your code bases and how do I defend my perimeter? How do I invest in a validation layer so that I can focus on the exploitable risk? And how do I invest in remediation? Because one of the stats I’ll give you is, uh, that in our platform, even as we’ve seen discovery rise, um, in the last year, backlog of critical vulnerabilities, uh, while our mean time to remediation has dropped by 50%, the backlog

Cal Al-Dhubaib: the backlog

Nidhi Aggarwal: of critical vulnerabilities has grown 29x. Not 29%,

Cal Al-Dhubaib: backlog,

Nidhi Aggarwal: That is the backlog, and that’s exposure debt. That is what organizations are sitting on. So if we don’t get to remediation, if we don’t fix these vulnerabilities, then we’re just admiring the problem. We found it. We’re admiring

Cal Al-Dhubaib: We’re problem aware. Yeah. We’re

Nidhi Aggarwal: Yeah.

Cal Al-Dhubaib: We’re moving [00:19:00] forward. Yes.

Nidhi Aggarwal: We’re problem aware. This is moving forward.

Yes. But we’re not fixing it, so found is not fixed.

Cal Al-Dhubaib: Might stay the same. And to go back to, to one of our earlier points is the idea that we can automate this all away, that- that’s just not there, but, but it’s also within reach. And I, I, I couldn’t say it better than you earlier, be brilliant at the basics, right?

Yes. Right. All of this is in pursuit of being brilliant at the basics.

Nidhi Aggarwal: the basics. It is, absolutely, because the basics are the same. We want to make sure that our, uh, right, our systems are protected, our data is protected, and, uh, right, our, uh, business is resilient. But one thing I will say is that, uh, the, uh, risk, uh, that we have to also manage is that, uh, earlier it was like, what can AI...

What model, um, right, how can I, uh, prompt inject the model to what does the model have access to? To now the risk is, uh, what can, uh, [00:20:00] AI do, right? What action can it take? And that risk needs to be managed in very differently in the sense of, uh, uh, how can we allow AI to be as autonomous as possible and take action while defending it to the level of autonomy- Yeah

we are comfortable giving it.

Cal Al-Dhubaib: And so that’s gonna take some new muscle that we have to build. Yes. Well, well, Nidhi, this has been absolutely insightful.

For our listeners, what’s the best way to follow your work and learn more?

Nidhi Aggarwal: Uh, first of all, thank you for having me. It has been a wonderful conversation for me, and I always learn more through your questions Cal.

Uh, the best way is to follow me on LinkedIn- Amazing ... and Twitter, and so that’s where I can be found.

Cal Al-Dhubaib: All right. Fantastic. Well, this has been incredible. Um, thanks again for joining the show.

Nidhi Aggarwal: Thank you so much for having me.

Guests

Industry Expert
Hosts
Cal Al-Dhubaib is a globally recognized data scientist, entrepreneur, and IAPP Certified AI Governance Professional advancing the practice of Trust Engineering for AI. His work focuses on how organizations design, deploy, monitor, and govern AI systems at scale. He leads emerging tech thought leadership at Rubrik, where he focuses on AI, cyber resilience, and enterprise trust. He is an exited founder of Pandata, an AI design and development firm serving heavily regulated industries and an international speaker and thought leader.