EP277: CISO as CFO, From Citi to Celery, It's All about the Cabbage

Security.

Hi everyone, welcome to the Cloud Security Podcast by Google, coming to you from Next 2026 in the Security Hub. My name is Tim Peacock. I'm a group PM here at Google along with Anton Chuvakin, senior staff in Google Cloud's Office of the CISO. We are the hosts of the Cloud Security Podcast by Google. We're joined today by Arvind Bansal, good friend, CISO at CNS Wholesale Grocers. Super stoked to have him on the show. Grocery stores are near and dear to my heart, not because I like to eat a lot, which I do, but because my first job, long before I made a living podcasting, was actually in a grocery store in Mansfield, New Jersey.

Shout out to ShopRite.

Wrote my college admissions essay about it. Worked out great. So tell us about the grocery business. And what the threat model looks like. I assume we're worried about, you know, somebody stealing all the tomatoes. Is that right?

Of course, yeah. So, you know, we are a CNS Wholesale Grocers. We are a $30 billion annual revenue company with 30,000 people, and we have 10,000+ retail stores as our customers. We also own 600 retail stores and warehouses across the US and say 106-year-old legacy. So you can imagine it's not just about tomatoes if we survive for that long, right? So, so it's about, you know, the top 3 things about our business in the retail. The first one is food safety. We all remember the COVID time. If things doesn't move, the food is not getting fresher sitting there. You want to make sure it stays safe. The second one is about supply chain. How do we make sure if things go down, the trucks keep on running?

This is physical supply chain, not software supply chain, correct? Exactly.

Yes, you're right. So, you know, AI is not going to move the food from one location to other location, right?

If it's an OAM, it might.

There's time to come, right? That's fair. Exactly. So really, it's all about how do we make sure the food is safe, the supply chain is intact, and lastly, when you go to that retail store, how do you make sure the shelves are always stocked with what you need? Right.

So I hope everybody here agrees this is not dry goods. This is actually a meaty story with beefy adversaries. Sorry, I really, I wanted to get some jokes in today. It's very important to me.

You're number one on the pun ratings. Yes, yes, yes.

So Anton, you have question number 2, please.

Yeah, I mean, I want to just pursue this a little further because it's— for some people, it may seem like it's more of a mainstream, mundane business, trucks, food. But you do use analytics. I mean, I know in the CPG industry, ML has a history. Deeper analytics has history. And now your business probably uses AI. At least your employees do and your business does. So given that most other conversations about security and AI come from people who are very well-funded, well-resourced, huge teams, projects with many zeros. How do you think about securing AI and securing ML, presumably for your type of industry?

We start at the business level, right? So if you look at retail business, the margins are razor thin. So right then and there, AI becomes the core of our business model because it drives those efficiencies. Now you bring the same efficiencies into the technology part.

Yeah.

How do I do better inventory? You know, we have millions of items in our warehouses. How do we place them at the right location so that the trucks are taking them with the least number of miles that they need to follow? And then how do we bring promotions and pricing to you in the right season? And how do we buy the goods from our vendors and suppliers and thousands of them at the right price on the right time. So AI is super crucial to how we do our business more efficiently.

So I see how you're using AI to move the business. How are you using AI to defend the company? 'Cause you have a background in the financial industry where there were gigantic security budgets. So what does it look like in the grocery industry to use AI for security?

You know what, financial industry, you have in some cases, unlimited budgets. And it reminds me of the moment when JPMorgan had a breach. Jamie Dimon came out and said, hey, here's the $500 million, go protect the bank. And after that, every security vendor we talked to, oh, we are great because we are at JPMC. We are great because we are at JPMC. And then we talk internally, how the hell do you deploy 500 solutions at a legacy bank from last however many years they've been surviving, right? So that's how the financial world— money alone cannot solve the security problem. And I personally believe, working in retail, you're not only a CISO, you are a CFO, because you're looking at every dollar. What is the ROI? What is the attack that I'm preventing to make sure the shelves are always stocked. And then you are also a CEO. You're looking at your balance sheet. What kind of resources I have? You're looking at scale. How do I drive it? You are making the decision around build versus buy versus outsource, right?

So in that world, as the CFO and the CEO, where are you seeing ROI on AI within your security group?

So there are 3 areas that we are looking at AI in our security group. The first one is, you know, heart of any security program is detecting the attacks, protecting the company from there, and remediate immediately. So that's security operations. And traditionally, security operations are inundated with thousands and thousands and thousands of alerts. And then you have fragmented security solutions deployment. So you have threat intel being done by a team and then you have your SOC teams coming together. So how do you bring those 2 pieces together? How do you bridge that gap? So AI, number one, it is helping us in SOC to bring that telemetry together and make our analysts more efficient. And the efficiency level that we have seen is ranging from 30% to 60%.

So, you measure it very carefully given that your wallet is limited. So, it's probably like wisdom for measuring the value of security would come from you rather than from bank where adding a couple of zeros to certain budgets may not be noticed. Well, I'm exaggerating massively, but you probably have better metrics for value than the well-funded companies. It's kind of an insight for me.

That is interesting. Not to mimic my friends here, but it doesn't sound like you're getting 10x more efficient. It sounds like you're getting 3x more efficient.

Exactly.

Yeah. So one is the SOC area. The other is, of course, the third-party risk management area. You know, the traditional ways of writing the questionnaire, getting the response, analyzing. I think those things are thinning away. And the third part is the continuous exposure management. So instead of hiring a pen test team once a year, now we have pen testing being done continuously. And not only that, finding where people can come inside our network and impact our system first. So that reduces the amount of work we focus on vulnerability management.

And any efficiency measures for the second and third? Because I feel like for the questionnaires, it's just nasty business. And if somebody takes it off my hands, I see value. But this is me being qualitative. Any, any numbers, any quantitative metrics?

Yeah. So there are 2 parts over there. One is number of vendors that we could cover through the questionnaire.

With the same team?

With the same team. So that number has improved by around 150%. You know, versus somebody manually sending, reviewing, and doing it. The second part is analyzing those questionnaires. So earlier somebody is looking and coming out. We have reduced the time by around 90%. Once you get a response, we don't look at it immediately. We run it by AI. What are the critical areas? And then focus only on the high-risk questionnaires coming out. Now, the continuous exposure part, in terms of numbers, I think it's hard to quantify.

Yeah.

But if you look at doing the test once a year, versus continuous testing all the time, that itself is increasing the safety of the organization.

Reduction number.

By thousands.

Yeah. Let me ask you a strange off-the-script kind of question.

Oh, Anton, you would never do that.

I would never do that normally out of 270 episodes we aired. But in this case, you mentioned the AI being used to analyze the questionnaires, right? Do you have this fear in the back of your mind that on the vendor side, AI is filling the questionnaires?

You know, the question is, should I be fearful?

No.

Or should I be aware?

I think, yeah, I think aware is more important.

So exactly.

So there are 2 parts to it. Let's just unpack it, right? One is, hey, is AI filling their questionnaire? Will the response be wrong? But if you take one step up, why do we do questionnaires in the first place? Why do we need those responses? You know, one is the risk reduction from legal and contractual liability. If I'm writing something and giving to other company with whom I'm doing billions of dollars of business, yep, hey, I'm okay with it, right? We ran security by contracts anyway for a number of years. The second part for that is, am I getting the right responses? Am I getting the responses in a timely manner? That metrics has changed dramatically. You know, the organization filling from AI forms.

So essentially, to extract a thought from this, if they use LLM and it hallucinates and they submit the wrong results to you, your AI may or may not notice it. Ultimately, if something truly blows up, it's their fault. Very clearly, a human on their side would pay.

Right.

So the machine—

so there are 2 parts here.

And, and I love the idea of Anton trying to like serve collection notice to Anthropic someday. That's great.

You know, you're taking me— I had another discussion about agents in retail. For example, Amazon has 25,000+ agents running in their environment today. So one way to look at agents is, hey, it's just another human for me. Every agent is tied to a human.

Yes.

So now if the agent is hallucinating and giving me a wrong response, who Who's accountable for that?

You'll grab that human and shake them until they pay. That's a good bag. Very good idea. Excellent. Excellent thinking. And you can't punish robots. That makes sense.

There's no way you can punish robots.

So you find a human whose ID is connected and you punish them. Lovely. I want that.

You know, another fun fact is that you and I, we have a manager who's watching what we are doing. And if you're doing something wrong, they can catch us, right?

Yes.

What about agents? How do you, the manager of an agent, know whether it is doing, whether it is drifting, whether it is hallucinating or not?

I mean, agent identity has been a famous unsolved problem, as we know, and people are solving it. We are solving it at Google. Other vendors are solving it. So I feel like this accountability argument strongly relies on it being solved.

Exactly.

No orphan agents, no agents left by departed employees.

Yeah.

It becomes very critical. Okay, so we are departing from the track.

We are way off. Also, you missed the very obvious answer. In the prompt, say, make no mistakes. Very easy to solve that problem.

Exactly.

So Anton, I think that the next question is yours, right?

Yeah, it is, but it's kind of a resilience question. And I feel like you described a business where you start to rely on AI. And lately, resilience for AI has been a lot on the mind, both with models degrading Access being not available, compute being overrun. Obviously not here at Google, but at some other places it has happened already. So how are you thinking of— well, if you introduce GenAI in your work streams and your workflows with thin margins, you allocate humans to other tasks and then something fails. How are you thinking about resilience to AI failures?

That's a hard question.

Thank you.

I'll say.

It applies. Yes.

You know, you have technology in your hand. You have technology in hand. What is your resilience if this goes down or this goes down? Printouts, water, paper.

That's correct.

So we have a 3-level of resilience where the first one is, of course, agents are doing the job, but we also have a period where current applications were doing those jobs. How were we taking orders before? How were we allocating inventory before? How were we purchasing before? So that makes the second layer of resilience. We can fall back to that. And the third one is, of course, you have to go back to manual. It's going to have an impact. It's going to have an impact on how we are able to stock the shelves, serve our customer, but we will keep the trucks rolling.

You'll degrade, but you won't die. That's basically the answer.

Yes.

Okay.

That's— and neither will the customers. This is actually really good. And I feel like there's a big lesson in this because sometimes I hear people to make calls to like become an AI native. And then resilience becomes a more critical issue because presumably if you're AI native, you can't easily switch to pre-AI way of doing things.

Right.

At least that's in the world. Again, I don't want to introduce more. It is a little scary. I mean, it's a little scary. I want to be AI native, but with resilience, to me, would be my take. Okay, Tim, go ask.

What's your resilience approach in general? Is it everywhere the same layered approach? How does cloud play into that? We are at a cloud conference, so I'd love if we could talk about cloud a little.

Of course, happy to. So I'll, I'll just go 3 areas in the resilience part and in the context of cloud conference, Google Cloud. The first one is, of course, what's critical for us? You know, how do we keep the orders coming, the goods rolling, the promotions keep going out? So that part we isolate to the high-risk applications, systems, processes, technologies first. And then as part of that isolation, we build those high-risk systems with resiliency. So the second layer is, you know, it's always, do I have immutable backups and am I able to recover from them? But more importantly, the human aspect of it. Do I have the right people to bring the application back? Will they be able to integrate it back?

Yeah.

Do I have the business people connected with my resiliency folks and the technology folks to bring it all together?

Yep.

The third one is, I think Google Cloud, there are 2 huge advantages of using Google Cloud. You know, the first one is the multi-region availability. So if something is going wrong in one part of the world, we know that we are covered from that perspective. And the second part is the innovation resiliency, right? So we talked about Google SecOps looking at it. If somebody's running a traditional SIEM, The AI innovation versus a Google secops with AI innovation. There's a huge difference how we move in that part. So resiliency becomes part of our conversation with Google Next, Google Cloud from there.

And you get it almost by default by using Google Cloud.

Hopefully keeps going higher, right?

Let's hope.

So before we start closing, because we have about 3 minutes, I kind of have a possibly sensitive/political question, and Tim was coaching me in how to ask it so that I don't annoy people. And I think I feel like properly coached.

You guys, you realize how bad Anton must be if I'm the one coaching him on how to ask a politically sensitive question? That is quite a moment to get through right there.

It's real though. So you worked for Citigroup, right?

Yes.

Where your budgets had many more zeros, your teams had many more heads, you had a lot more systems, a lot more diversity of what you could— resources you can summon. including what you can summon in an emergency. And now you are at a company with much thinner margins and smaller budgets. How is your thinking as a security leader changed? Like, how have you adapted and what changed? And maybe specifically, what changed did you build by outsource?

And maybe one follow-up to that is when you were making that career move to CNS, was that part of your thinking is what would be fun about the role or what would be scary about the role or maybe both?

It's the scary part, you know? What would be the scary in that role? I'll tell you what my take is. You become one of the best CISO if you're working for a retail business. And I'll tell you the reason for that. You know, if you're working for large financial with big budget, yes, you can muster everything in the world, buy the best technology, get as many people you need, start deploying it, make sure you're protecting. But here in the retail business, you're not just thinking about technology. You are thinking about business like a CEO, like a CFO. Every dollar matters. So that shifts your thinking about looking at product, looking at capabilities, and looking at deployment of those parts. So in my opinion, I think it's great to be a financial CISO, but it's even greater or better to be a retail CISO. So that you learn the business side of it.

I like that answer. There are people here who know me that I can't enjoy hobbies unless they scare me and have a risk of death. And so that same, you know, being scared of the thin margin but sharpening your focus, I really relate to that. So I like that a lot. We have actually just gotten bonus time on the clock, Anton.

Somebody has bumped up our time.

It's like in a video game, you get a bonus life, I think. So I get to ask a bonus question, which is around IoT and OT convergence. So you've got warehouses where things will literally melt if it goes wrong. So how are you bringing that together? How are you monitoring that? How are you keeping that secure?

Yeah, I'll give you the overall vision and the context, and then we can come to it. So in the world, we have 8 billion people, and there are only about 25 million people who are writing the software codes today. That's like 0.6% of the entire population dictating how 99.4% of the people in the world will use software. Now imagine the AI is taking you where everybody can write the code. And if that percentage of AI writing the codes increases even up to 5%, 6%, Of the world, what is the next constraint for the attackers or the bad people? It's going to be hardware. And so every forklift, every pick-pack-ship scanners we are using, they will become target. And that's the next wave of attacks. I feel we have already faced acid forklifts.

That's disturbing.

Oh my God. I thought it was the funniest thing I've heard about the internet-connected locomotives, but internet-connected forklifts is terrifying.

Look, technology is always connected, right? So, so now the next attack surface is going to come from software into the hardware. That's where the focus is going to shift. So I believe IT and OT, they're going to merge very rapidly. I was just talking to somebody this afternoon. They asked me the question about, hey, what's the AI benefit? And I'm like, my wife cannot see any benefit from AI. You know, she has a comment, can AI do my dishwashing? The answer is no. So it made no difference into my life. Is AI going to do my laundry and folding the clothes? No. So it made no difference in my life. So it's like, those are the factors I think Where the world is coming next, they're going to merge. The security of that is going to be elevated because that's where the more attacks are going to be seen.

So I hate to do it, but we're at the end. We have to ask our traditional closing questions. So one, do you have a tip to help people out there think about adopting AI better? And then 2, do you have recommended reading? And it could be anything, just not Anton's blog.

Let me ask you for a bonus tip. Again, best advice for being a CISO of a more constrained organization after being a much larger, much more financed and well-funded organization.

So, AI and consent.

Yeah, I feel like this is very useful for many because, again, I know that's a concern, right?

So, 2 tips and 1 reading advice. All right.

All right.

The first tip I would give is, yeah, I was just having this discussion with another CISO about using agentic AI in security operations. So he started working on agentic AI back in August last year. He said, Arvind, it took me 4 months just to get the tier 1 work done. And the reason, root cause for that was the people and the culture. Are they accepting of agentic AI? Are they accepting of the new innovation you're getting? So the tip I have for you is don't worry about technology. Focus on technology, focus on people in your organization. Who are the key people with the knowledge? Who are the people who will become evangelists of agentic AI? And then try to get them an environment to focus on that.

Sorry, are you telling me that there's an organization out there where there were people in the SOC who didn't want the agentic triage?

Yes.

They wanted the key triage?

It came up today in the—

They wanted more alert fatigue.

No champion, no AI was her take as well. Oh, yes. Very real.

So you're telling me, oh man, Stockholm syndrome big time. Okay. So you just blew my mind.

Yeah.

It's going to increase. So that's the first part I would say. The second part is embracing it, you know, not just at your work, but what can you do with AI at home? How do you integrate things? You know, somebody was giving me a tip. Hey, I just created an agent which looks at all my email, sends a message to my kids about what they're going to do all day, sends a message to my nanny what's going to happen. And I was like, oh my God, I wasn't even thinking in those directions using agents for. But that's the kind of creative thinking we need to bring from our personal life and then move it to the business life. So those are the 2 tips. People culture first. And then second one, our own thought process. How do we use in personal life to build that skills of using in the professional life?

Do you have recommended reading? Could be anything from Mary Shelley to Mary Oliver.

You know, I do a daily AI podcast. So Daily News AI podcast in Apple News that keeps me really, really sharp. So I would say just go to Apple Podcasts, search for Daily AI News. It's a 15-minute podcast, but it gives you a lot more information about the AI world, what's happening. But my favorite podcast is, of course, you'll say it.

There you go. There you go.

Do it, yeah.

You cannot miss those 200+ podcasts, guys. I learn so much from what's happening and the discussion. Those are the real deals. Yeah.

You are too kind. Thank you very much for joining us today. And thank you all for joining us live here at Google Cloud Next.

All right. Thank you. Thank you.

And now we are out of time. Thank you very much for listening and, of course, for subscribing. Please subscribe so you can get new episodes piping hot. And if you love our content, please drop us a review on your platform of choice.

You can find this podcast on YouTube, Apple Podcasts, Spotify, or wherever you get your podcasts. Also, you can find us on our website, cloud.withgoogle.com/cloudsecurity/podcasts. You can argue with us on the Google Cloud Security Community site, googlecloudcommunity.com.

You can follow us on X, x.com/cloudsecpodcast. Tweet at us, email us, argue with us. And if we like or hate what we hear, we can invite you to the next episode.

See you on the next episode of the Cloud Security Podcast by Google.