Name: Securing APAC Financial Services: Building Resilience in a Dynamic Threat Environment
Uploaded: 2026-08-11T06:26:11.975Z
Duration: 1 h 6 min 36 s
Description: Securing APAC Financial Services: Building Resilience in a Dynamic Threat Environment

Transcript for "Securing APAC Financial Services: Building Resilience in a Dynamic Threat Environment": Okay. I think we are live now. Hello, everyone, and welcome to today's webinar. My name is Vinay, and I will be helping support today's live session. We are so excited to have you all join us. But before we end the webinar, a little bit of housekeeping. On the top on the top right corner of your screen, you will see the live chat option that you can use to interact with your fellow attendees. In the doc section that is located next to the live chat, we have attached a few documents that are relevant to today's topic. In case you peep if in case you wish to read more or learn more about what we what's being discussed, you can just click on them, and you can download it right away. On the extreme right corner, you will also find the q and a box. We encourage all of you to leave your questions in that q and a box as we progress through this webinar. We'll make sure to address all your questions towards the end of this present presentation. Quick reminder, this session is being recorded, and we will share the recording with you within forty eight hours. Now please join me in welcoming today's speaker, mister Shati Kumar. Hey. Hi. Thanks, Vinay. Hello, K. everyone. Thanks a lot for that. I see we have got 22 identity. Mhmm. I shall we start with now, or let's give it another minute? Because I saw 40, and now I see 22. Yeah. I think we can wait for a couple more seconds, and then you can kick off shortly. Yeah. Let's wait for the, one zero two my time. So few seconds, and then. we'll kick off. Alright. Till then, I'll just, start engaging with the audience here on live chat, and. I'll let you, I'll let you go. Thanks, for that. Shruti. I'll take it from here. Thanks a lot. Yeah. It's about time. Okay. Thanks, everyone, again, who have joined late. Quick housekeeping for them again. I know it might be a repetition for others, but, do make use of chat. We want it to be interactive session, so do drop in your comments there. There is a doc section to download some of the material, which I will prompt as we go along. I will point you to those documents where they are relevant to the topic, and we'll go down from that direction. Q and a as it suggest. If you have questions, feel free to, you know, write your comments in the q and a section, and we'll pick it up from there as we go along. There is a q and a section in our presentation towards the end. Right? Let's give it a start. A lot of people have joined. So first off, let me officially now open up. Thanks a lot, everyone, for taking their time out. I know this is a busy schedule for everybody, but, it means a lot for us to have you on this session and talk about what we are doing. Right? So today's session is really simple, in terms of direction we are going. I want to talk about banking and financial services, specifically in APAC and how CloudFlare has been helping some of the banks, some of the financial organizations, some of the payment gateways, for example, in crypto to attain better governance with AI. Right? While AI has been, I would say, flying quite at a hyper speed, I would say we all would agree with that. Governance had been something that was not at the front seat. So that is what I want to address there. Right? Quick introduction, myself. Yeah. Thanks for that. Yarl, let the emojis keep coming so I know that I am on the same page as you. Okay. So I'm Shakti. I'm a principal solutions architect based out of Cloudflare's Singapore office. And my charter is really to help with banking, financials, and payment gateways. Basically, BFSI, segment industry here, specifically to help help them with monetizing on cloud or CloudFlare related products. Right? With the background, I've been in industry for twenty years, and I've been in different countries within APAC, currently in Singapore. Been with AWS, Microsoft previously, where I've been more or less dealing with Standard Chartered Bank, BBS, now recently with crypto, in India as well, Standard Chart sorry. State Bank of India, CB, if you have heard about those. So, yeah, I've been there. I've helped customers with their, regulatory framework, maintaining a regulation year after year, maintaining it, as we all know, is hard to do. And not so great news is it is getting harder as we get AI into the mix. But fret not, we do have, AI on our side as well. If it is there on the other side, on the bad vector, on the good vector as well, we can use AI to defend ourself and something that I will cover as well. Right? With that, let's get cracking. Let's see. Okay. I always want to start with why. Why now? Why are we talking about this topic? Right? So, what has happened is the world is moving fast as we all know. From evaluations, prototypes, testing to production, simply put. Back in 2024, it all started with ChargeGPT's, you know, simple test on chat where you would ask it a simple question. It would do research for you and come back, and we all were amazed. Okay. It has a lot of potential back then. Right? It all started from there where we were not really thinking about bringing it to production. And then, of course, governance, which is the prime topic here, was not something on the forefront. It was a two page PDF document. Like the old good days, regulators were okay with the yearly PDF report going down, and then SOCs and such were pretty okay. Fast forward to now, 2026. Right? We are looking at two thirds of banks' back end processes, like loan decisions, like KYC, like fraud scoring and equivalent being already taken over by AI in production in some form. So I'm talking about frontier models as well as open source models already gone into production in some form. May not be completely taken off, but then some part of it being automated or such. Now that is one. Two, employees are using frontier model as well as open source models again, Chargebee Ts of the World, Anthropic, Gemini, and such. Now all that is correct, but then, governance was not something that has kept up. Governance typically is something that comes from regulators, and I'll connect the story back there. That is still sticking to 2022, not 2024. Why? Because nobody was planning for it. That makes regulators as well nervous. There is a lot of development that has happened in the last, let's say, about eight, nine months. I'll cover that as fast as I can. I'll I'll showcase what regulators have come up recently. But the gap, while regulators are now talking about guidance, not a law currently. Yes. There is a mix of it, but then because AI is new, they are trying to be having a soft approach as we say. So they're asking the right questions for now. Where is your AI data for the resiliency, of course? Where is all, sovereignty, basically. Right? And then how you take a AI decision? What is your third party discipline? These are the fundamental questions that they are starting to ask. And good or bad, lot of large organization, I'm talking about enterprise grade, financial organizations, do not have a clear answers to that. With the agent API, that becomes even harder. So this session is really to address that gap. How do you answer, compliance? How do you answer regulators? How do you keep up with the changing regulation landscape without having to break your bank? We need to say without having to pay a lot, runaway cost from there and, you know, making the budget impossible to achieve. I've been on both sides where we met the compliance, but towards the end, it was already about 30 x the cost that was previously estimated, and then the project had to be shared. So so not having to go there as well. So that's where we are, and that's what I'm trying to address. If you have any questions, concern, if this is not what you wanted to talk about, feel free to, drop a message on the chat. We will cover that as we go. But this is where the topic is right now. Now quickly having a slide which talks about the same problem, but at a little bit more depth. This is a full slide, and I'll introduce to you what that means in a minute. But then AI governance, as soon as we hear, we all have been most of us, I believe, would have been aware that when when we hear governance, let's be bold about it. Right? It sounds boring, and it sounds like not so interesting task to deal with. We all want to be on the forefront of leveraging AI and solve business challenges and innovate with it. Nobody wants to be, like, you know, create a defensive mechanism and figure out how do you we piece together some in incident and then create a response for regulator. Nobody wants to do that. But the truth is that is also a job we need to take care of and job we need to take care of very well. So this is where governance is not by definition, it's not just security. It is compliance, cost, and operation. I for those who don't know about it, let me introduce all of them. Security, jailbreaking, PII, prompt injection attacks, model exfiltration. Meaning to say model was trained on some data, somebody is trying to retrieve the data or PII in that data, that is also a security breach. Shadow AIs mean to say those where employees built some kind of scaffolding for AI and they forgot. Or worst, somebody is trying to make use of tokens that you paid for, and then that is a shadow eye. How do we deal with that? That is coming under security. AgenticEye takes it one step forward. I have a slide on specifically on that, but then yeah. The older ecosystem, older narrative were designed for humans. The whole, poster security poster were designed for human, the IDPs, the ABACs, the RBACs. Right, attribute based or role based was designed for humans. With agentic AI, it is pushing the boundary. It's testing the limits. It's going beyond. It's breaking. If if few of us are around and have read the news, look at what, hugging phase happened just about, I think, one or two weeks ago, where OpenAI, break away from a sandbox, and then it was able to breach into some of the, not so allowed areas, and it was able to expose credentials. And it was not a targeted attack by a human. It was completely agent going rogue. So that is not one scary, but then that's not one. There are multiple of those that are coming up now. Right? So So how do we deal with that? Being a bank, we cannot simply, think that that will not happen to our bank. We have to prepare for it. Compliance, of course, that that's that's the job. They they are looking at it, and they keep in sync with each other. Look at DBDP from India, for example, or MAS from Singapore, HKMA. All of them now have some sort of guidance slash, come mandatory wordings coming out, which has AI in them. Yeah. Restriction, cross border AI, explainability, risk of AI. Now they have started to come up with AI. And I I think the next slide actually takes it at a little bit more deeper level. Cost. So we do have, while we are meeting, as I was saying, the the the compliance risk and security risk, the cost is something we need to trade off. We do not want to have multiple copies of data, for example, the old problem. Right? We do not want to have all the world GPU sitting just because we wanted to infer one piece of request. So how do we mitigate that? And finally, operation, this is the bigger piece. This is what we have not been thinking. All of us have been testing, evaluating it in, let's say, a smaller spatter setup. When it goes to production, things are different. Dimensions are different. All of a sudden, vendor locking comes in, which was not something we were thinking of. Model drift comes in. Audit trail, which is very hard to keep up. I'll come to that as well how what is audit trail and what it means to piece together information in a live fashion, excuse me, that, regulators need. How do we get there? So that is a hard one, plus we need to comply with regulations on that as well. So I'll cover that in a slice. Sorry again. Now agent AI only makes it even harder because it is not just about MFA, but standards access controls for them has to be different. OAuth 2.0 is a testimony to that if you know about the subject there. Right? Now this slide is a poll question. If you see the green color on the top there, top right corner. The objective over here is really to see how it makes sense for you. There are no right or wrong, questions. This is more for me to understand the direction that I'm taking or or these risks are something that you are worried about, your organization is worried about something in next, let's say, three to six months, yeah, then please, click on your option and let us know that that resonates well with your approach there, right, or your thought process there. So I will ask Vinay if you can, kick start the pool so that my audience can click on it and then find out or give us their votes on around it. Let me see if we have pool here. Yeah. So everybody, if you are looking at your screen now, there is a poll section, and there is this after chat between chat and doc, you will see pool, and you can submit your responses there. I'll give it a minute. It's 01:14 my time. Let's say at 01:15. I'll take a pause, and, we'll see the results, hopefully. So giving you a minute. I do see some of the votes are coming along. For those who have not heard, or maybe, the audio was not clear for them, once again, this is a old slide you can give. There is no right or wrong. There is no competition. There is no accountability here. There is no risk, basically. I'm asking which of these risks or it it is possible all four of this these risks are something you're concerned with, are something that your organization is thinking about. Right? Just to keep it intact. Okay. Give another ten seconds. Okay. I think we have got okay. So security risk. When we say the prompt injections of the world of model envision, shadow AI happens to be number one problem that we are looking at. So we and closely following that. So I've got about, what, 11 votes to security, eight to compliance, and then a few to cost and operation. So thanks a lot, for those who have participated and given their thought process around it. I will keep my session more, centered around those two while, of course, cost is important. We don't want to have a runaway cost. It's just one or maybe two slides, but we'll we'll not spend too much of time on that. I'll try to spend more time on security and compliance on how the world is doing it. It's platform agnostic. My approach is not really to sell CloudFlare solution. Yes. That is my API in the end. I would not shy away from that, but then you can do it yourself as well. Approach is to show you a working model that has been there in the industry. And then, of course, how CloudFlare has been doing it. While you can do it yourself, you can, you know, find your own chosen vendor, favorite vendor, and you can do it from them with them as well. Alright. Let's close the you know, you can close the poll queue for that. Okay. Coming back to the slide. Let's see. Yeah. We have got the answer. Quick update here. I wanted to start with a topic where a little bit of development has happened. If you have heard about Mind Forge, give me a thumbs up. If not, it's okay. But quick idea. This is Monetary Authority of Singapore where 24 financial institutions came together, and they thought through what are the risks, what are the mitigations, and how do we get there. So what, why, where, how kind of questions. Right? I was lucky enough, in my previous role to be part of that project as well where I wrote a bit of, policy document. Basically, provided some of the inputs towards it, which eventually, obviously, became policy. It's not like I wrote the policy document, but then they did involve vendors, SIs, AWS, Azure, Google, Nvidia, for an example, plus 24 financial institutions, plus some SIs put together. They came up with a what is called as, handbook. Just for everybody's understanding and clarity, this handbook is available for download. This you can look at, the down the doc section, next to chat, and you can download it and you can read about it when you have free time. But very simple, they looked at areas like oversight. This is more or less like a human error or something which human cannot, you know, probably cover. Risks, life cycle, and enablers. So these are the operational risks that they went into operational areas they went into. And they covered not just AI. They covered, generative AI. They covered agentic AI. They have covered all three phases of AI that are prominent today, of course. We do not have, artificial general intelligence or super intelligence yet, so not much of that work is here. But long story short, I don't want to bore you with, with the document which was probably written by somebody else, But there is something that is a key message there. Right? There is a guardrail. There is a there is a standard one liner, which is at the bottom if you if you're looking at my slide. Let me see if I can get a pen here. Yeah. I do. I do. So I'm talking about this section. Regulators are now not happy with just logs or rather PDF documents yearly basis. They now need much more. They now need, for example, evidence, a live evidence, not just logs, but live evidence. So so how do we get there? For an example, I'll give you example which could clarify. Right? So PII, which which we all know, like passport numbers or, let's say, NRIC number for Singapore or Aadhaar number for India, for instance, or HKID for Hong Kong. Now these are not supposed to be landing outside of the region. For example, if I'm in Singapore, it should not go and to The US. But then Anthropic, Bedrock, OpenAI, the latest and greatest models are in US. So they go there by form of tokenization and and encryption and so on and so forth, which may or may not be something that regulators like depending upon where you look at. But now they want an evidence that it's not about going there and doing something as a reactive approach from the LLM side, in The US side, for example. It must be able to stop in the region. My case, Singapore. And you must be able to prove it. That is much harder than previous versions where we will just had a PDF document to upload and things were okay. Now no more. So how do we do deal with that? I'll come to that. Right? I want to see if I have any questions. Taking a pause. Let me know if there are any questions on the internal chat we have. I'll stop there. Otherwise, yeah, I'll keep waiting. Okay. Okay. Thanks for that. No questions so far. Great. Okay. So, moving down to the next one, regulators, what they are asking. Now at the when you look at this slide, of course, this is a busy slide. There are a lot of wordings, and it could easily get confusing. And I'm not trying to solve every regulator's ask. It's impossible. I won't do that. But what is happening is there are common denominator. There are common asks. Everybody is asking for one simple thing, accountability with identity guy. The agent must be accountable with one credential. It can't have, for example, rules. It can't have, let's say, for example, STS alone, which which keeps changing. It should be accountable with one database or one ERP solution. It should be accountable with one network. The blast radio should be you know, those are the guidelines. Guidelines become law eventually, of course, as we know. So if you look at the the common ask between India, Singapore, Hong Kong, because majority of my audience are today from these, three countries when I was looking at the so I picked them picked these, regulations and build something around them to talk about. They all are asking the same thing. Where is your data? Can you explain your AI decisions? And can you prove controls are working? Which I was just speaking about that you have some PII. You do not want it to be transmitted, but can you prove it? How do you prove it? And then what what happens if somebody forcefully do it? Do you have mechanisms to see if somebody is trying to bypass it? Yeah. And there are even scary part where, let's say, if LLM was somehow, you know, compromised and it wrote a code, which agent needs to go and run, which is malicious. How do we defend that? Have we tested on what are its capabilities of going far? All these are problems written. Fret not if you have not thought about it. We have We are a security company, so we'll cover that for you. That is in the solution part, but this is very common. Now I want to, again, come down to one more poll question, poll slide. Again, I mean, there are not a lot of it. I just want to make sure that we are going in the right direction. It's not something I I keep happening. I mean, I'll keep doing every other slide or such, but, probably this is the last one where I want to pick up a direction because we know security and compliance were the previous selected topics. Now regulators, at the very least, the fundamental questions that they are asking are these four. Number one, which AI models now? Of course, you can find from logs today that which AI models they are using, but the better part is what I I'm more keen is those which are not even approved by your organization, where staff is using, for example, personal account. Can you see them? Do you have the ability to see if somebody has been using browser or API or SDK or some mechanism of using some kind of model from their office laptop, for example. That's one. Second one, do we know if the data so this is more a data sovereignty question. If the data was within the jurisdiction of your country, could be RBI, MA, S, H, K, MA, wherever, or did it fly out? How do you know, and where is it logged? How how much time does it take for you to reconstruct that information is also a question at the back. PII, if somebody wants to yeah. The guidance is do not put your sensitive information, but if somebody wants to forcefully do it, like, this keep spaces in between or they do some kind of, you know, thought process around it and try to bypass the system and enter PII. Is this something that can be stopped? And the last one is one six one eighty days, which is more of a d p d p, but I want I'm I really like the approach which DBDP has done for SIM consistent logging, where can we play back an incident which happened, let's say, six months ago, one or three days ago, and still reconstruct what happened, where it started, who processed it. And from there, the flowchart wise, who acted on it, where it was locked step by step. Can we do that even after, let's say, or let's say at one eighty day mark. Right? Not beyond one eighty, but at one eighty day mark. So I want to take, again, I will ask, Vinay, request here to raise the poll. Again, everybody, all they got to do is look at the poll questions and select which of these you can answer confidently today. And, again, no right and wrong. This is not accountability. This is something I want to make sure that we all, we all know where we are heading, the risks that we know. At least knowing the risk is 50% of the job done, as we say, then we can work around it. Right? So these are the four that regulators are asking. I've been in these calls where regulators come and ask questions, and they kind of ask in a law or let's say, you know, where you have to provide word by word in a in a manner that is code of law kind of a language, not like. So so, I'm asking in a very simple English there, but, yeah, if you are able to answer those in official response standard that regulators want. So let's see. What time? 01:27. I'll give it a minute to see if, any of those answers are there. And while the votes are happening, I'm also looking at the question. Thanks for that queue, Vinay. I need to expand the screen. I see one question right now. Okay. So I do have a question there, which is how do runtime guardrails work? I see that. Thanks for that question. I am going to cover that as we go along. There is one slide that talks about it. I think another two, three slides later, I'll cover that question. So, if any other questions, do let me know. Okay. Let's take a pause. Here, I do have some of the words. Most of us are able to answer either one or two questions out of it, and which is pretty standard. If we look at at and no no shame in that. I mean, even if I'm working for the bank, for example, the banks where I've spent less than six years, eight years at a stretch, I would not be able to answer because these are complex. Things are changing. Banks or financial organizations or even payment gateways are not designed to answer these questions so easily. That's a complex machinery out there. So how do we fix it? Let's now talk about the solution piece. I've been talking about the problem so far. Let me bring down to solution now. This is where I was talking about platform agnostic. There's something you can do with any vendor of your choice. Maybe you might need to do a lot more work, but it's not, something that must be done from Cloudflare. I'll bring to Cloudflare in, I think next one or two slides, and we'll talk about that. Right? So first off, there are three pillars, and this is at the back of, of course, the research that other companies have done, and then they have bought, their solutions around it. There are three pillars to solution, visibility, control, and auditability. Now visibility means you first see what happened, what happened in your organization, what's happening in your organization, then you can control it. The narrative I have mentioned there, you cannot control or you cannot govern what you can't even see. So what happens? Invisibilities, you should have a full inventory of every LLM, every request, every response in real time mapped to either team or the request a person or by the model, by the region so so that you can answer those questions that regulators are asking. Right? This is accountability basics. Now this is the core. This is where a lot of tools are available, which might be able to do a great job or maybe able to provide you with some level of accountability and visibility, but not everything. That's one. Once you have visibility, next step is naturally you put some control. When you know something is not right, you try to, you know, let's say, rate limited or let's say, put a token budget or let's say, you want to work on blocking PII, for example. Right? That's control enforcement, policy enforcement. And auditability is where you now want to have logs which shows that what was blocked, what was approved in SIM format. SIM, for those who don't know, this is the format which is required for a lot of regulators as well as, MTTR, which is mean time to resolution as well, where no, end to end log in there, plus you do not have any credentials or any sensitive information. That is masked. It is kind of metadata for those kind of sensitive data, but then that log is required by regulators as well. Right? So simply put I'm not going to go, too much detail into it because this is philosophy end of the day. But, yes, this is where our solution cloud based solution is also designed. Right? And what is the cloud based solution? Quickly coming to that point as well. Okay. So those who probably are paying attention, thanks a lot. Those who are not, this is one of the important slide. Please pay attention to this one. We call it as AI gateway. If you are hearing for the first time, give me a thumbs up. If, you have heard about it, you don't have to respond to that because then we would know already. Right? For those who are hearing for the first time, do give a thumbs up. Thanks a lot. Yeah. I appreciate you, clicking on that button. Okay. So for those who have never heard about it, first off, good news. This is a low barrier, solution from CloudFlare to start. This is not something which typically should be costing you. You know? When I when I say typically, I mean to say things like analytics, which I'll cover, analytics, caching, rate limiting, DLP scanning. If you heard about this term, great. If you have not, never mind. I'll cover that. I'm going to be free for all plans. Only when you have advanced things like persistent logs, if you want AI on those logs and reconstruct something else, yes, they might be cost involved. But then starting off with AI gateway is free. It's as simple as that. Right? So what does does it do? As I recently covered, visibility. Maybe say we will give you which LMS are approved and which are shadow. Somebody is using without approval, for an example. Right? Which teams are consuming the most tokens? These are simple questions. Very very hard to do when you have agent decay, when you have shared credentials, when there are teams who are changing their credentials all the time, when there is no tagging between the teams. How do you get there? So there's a science behind it, and I'll talk about it later if if you want to go in-depth. But we give you visibility so you know which models, which region, where is your data, where it is processed, what is allowed, what is not allowed, and such. Once you have the visibility, of course, we can give you some of the reports which gives you, you know, the data, you know, structure format. So how much time does it take for you to get this kind of visibility? One week. It's not hard. We use what is called a secured web gateway, and we intercept your application traffic. I'll tell you I have a architecture slide which more explains if you want technology wise how it happens. Those who don't like to understand more technology on this, the tech, if I keep aside, we become your broker between your current infrastructure and you wherever it is going, wherever the flow is. We sit in between, and we are able to see and tag. Of course, there is a policy that you need to set and come out with report and inform you about all these which are something that regulators need. So that is point number one, visibility. Very easy. Point number two is really talking about control. Right? Now what is control? Control is where you want to put in, some kind of enforcement. For example, I was talking about just now, a prompt where somebody go and enter Aadhaar for India for those who don't know, analyzing for Singapore, HKID for Hong Kong forcefully. They should get back a response that this is stopped at the edge. This is not coming from LLM, which might be sitting somewhere, not from inferencing, not from GPU, not from OpenAI or Gemini. This is not coming from that. This happens right at the source. That is what we bring onto the table. Right? So even if PII was there, it was stopped even before it leave the system, leave the boundary itself. Right? Plus routing. Now this is very important. There are couple of things here. Let's say if you have PII or sensitive information that must be processed from specific country, in country, and you have a system, like, your own bring your own, on prem system. We work with that. By the way, we work with the architecture slide is coming up where I would explain. We work with your own, bring your own LoRa, models which you have trained or we work with Frontier models. We work with open source models. It's a big list that we work with, almost everything. Right? So we help you route the traffic to where you want it. If it has sensitive information, go to something else. If it has not sensitive information, go to something. If it is complex, go that routing. If something fails, I guarantee you always have to think about that as well. If, let's say, AWS will which has it has happened fifteen hours of outage happened on API side, for example. With context, how do you switch? That is also routing. We take care of that part. Then comes rate limiting where teams are experimenting. We all know agents can be fired up from nothing 2,000 in no time. How do you then prevent a runaway cost? How do you ensure that throttling happens, rate limiting happens, and so and so forth? And somebody select. Those are the kind of aspects we bring onto the again, this is not something you have to pay for. All this is included in most plans. Most paid plan already have this. Right? Now that is control. Pillar number three, if I can move my cursor, I think it okay. Got it. Auditability, this is more of the logging. This is more of where you have the evidence reconstruction of the event where let's say, if if a if a regulator wants to come back and ask you what happened in that specific day, how did you block, show me the evidence. Reconstructing back is something we, will be able to cover for you up to one eighty days or beyond that for that is configurable. And the best part, you do not have to pay for the log as well. I was saying advanced logs, if you use on Cloudflare, yes. But let's say if you have your own chosen s three based glue, I don't know, or you have some kind of Datadog or, other, you know, telemetry analyze analytics BI or such you want to use by your own because it is working in your system right now, we allow that. We'll push the lab logs, which are SI inconsistent, into your chosen vendor, and that will look at it. Right? Yeah. Those are the three pillars. I want to talk about agentic AI as well, and then maybe talk about the the, architecture slide before I take a pause and see if there are any other questions. I do remember. I do have one question that I needed to answer from previous time. I'll look at that in a minute. Now agent AI, just to put a perspective, I'm sure you would have heard about this term. If for those who have not, of course, these are autonomous systems which can go and act on behalf of you. IMDA, Infocom Media Development Authority. I picked up this name because a lot of people have joined from Singapore as well. Other countries are also building up something similar. Now this is not a law right now. This is more of a hard guidance, but it will eventually become law. So what are we talking about here? We need to have a lot more governance with agents because there is no human in the loop. And when money or or let's say data is involved, we need some kind of approval. So regulators are asking, who is the agent? How do you get that answered when you do not have a shared credential? When you are not somebody who is using rule based authentication, for example. You need to have specific ID, which has only one purpose, one transaction to happen on one agent. That is where you get the level of, accountability that regulators are asking. What it can do? When we are test have you tested it? So this is where, of course, we provide a solution for this. If you heard about hardness architecture and all, a very long concept. But simply put, you could, you know, take your own environment into, physically separate infrastructure, and you can test. For example, if you heard about Mitos glasswing or such, you can test those models on your own data and your own infrastructure so that you are not learning a hard lesson. And you if you got what I mean, you figure out your own bugs, you figure out your own weaknesses in your system before audio data or your posture security posture before a real attack happens. Right? Have you done that? How much of effort is required to do that? Can it be done? Those are the questions that you need to answer when regulators ask. This is also important. Part of that is is it behaving correctly? Can you approve it? If there is a traceability, if there is auditor and all. Now I'm going a lot into adverse, adversaries or I would say advocacy as well. This is not really talking about what can happen or what has happened is not reactive. This is more guidance towards protecting scenarios that can happen in future. So it might sound like, yeah, there that is a open ended situation. How do I deal with that? But, yeah, be with me. That is what right now we are dealing with because this is a new incubating technology. There is a lot that are that are going to change along the way. But for now, the biggest takeaway is having to have a mechanism, a governance system around agents. Right? So quickly switching down to the architecture slide because that answers a few questions here. Right? Now for those, who have been paying attention, thanks a lot. But those who do not haven't been, looking at CloudFlare previously or don't know about where it plays. Number one thing, there is no big bang change. There is no DNS or migration changes or such. So whatever you have in your Bedrock or Azure or your on prem stays at it as it is. We do not change that. Good news. Right? Is there a code change? I would start say no. There is a single change with what we call as endpoint on these systems. What are these systems? These are the applications which are already there, front end applications, APIs, custom facing, or internal applications. They need to point their endpoint URL to Cloudflare. That's the single change they do. As soon as you do that, Cloudflare AI gateway gets activated. It gives you visibility, which I showed you in the other slide. It gives you the eye detection, caching. And these things are more or less you don't have to pay for this. Right? Routing, log, of course, it is different if you store logs on us. Yeah. Different story. Rate limiting. These things come up. Okay. Something happened. Okay. These things come up from within a week or two, you see the benefits. Now this is where the agent comes in as well. We do integrate with your agent. Same control mechanism, same rate limiting, same routing mechanisms, same caching benefits, get there. I have I have a slide on caching as well. Right? So, last one is firewall for AI. For injection detection. Jailbreaking is also coming. So it's not like you have to typically, in a hyperscaler, you need to build all this together yourself. We give that as a package. It is inbound I mean, inbound, sorry, where we process all this in one sequence rather than having to go all the way. Now I want to take a pause and see if there are any further questions. And remember, there was a guardrail question. I have not got the slide yet. I'll get to that. Okay. I'm checking on questions. Not so far. Okay. Okay. Alright. K. Noted. Yeah. Team is quite, slow for today, I suppose. Last time we did a session like this, there were a lot of questions. Never mind. Let's see if this excites you or this, shows you this, this brings any questions in your thought process. Right? Hey, Ravi. Sorry. to interrupt you. We have actually received a lot of questions, Yes. Really? on the on the on the left hand side panel, the q and a. You will. see all the questions coming in. Sorry. Sorry. My bad. Okay. Okay. Okay. Oh, Yeah. that's that's. what I was thinking. Yeah. Yeah. Some okay. So I I'm reading the question one one by one. Let's see if I can answer those. Some agents pick skills dynamically. How do we secure such agents? Yeah. So agent security is a topic I'm not covering in detail there. So we, as CloudFlare, have a six pillar architecture for that. Now how do we do it? We run agent in a isolated sandbox with credentials which are inserted them in them on the fly. I mean, you said they are not in some secret store or not in their environment variable so that when memory dumps at the node. You know, when you are trying to take logs, the credential doesn't come up there. Maybe a bottom line to that is we deal agents in a highly secured, physically separate without any credentials in the memory kind of environment. One says, statement if I can make so that the agent physically cannot come out. Even if it breaks into the system, it does not come out of the sandbox and do the harm that we don't want it to do. Yeah? Where the logs are stored. Next question, for Manish. Thanks for that. Yeah. So logs are something that we capture from AI gateway, and we store it on your chosen, platform, s three, or we can push it to we have our own object storage for r two. We can send there a Datadog wherever. So either you store on CloudFlare or you can store it on anywhere. By the way, this platform is quite big. We have three thirty five cities in the world, one of the biggest in the infrastructure. India itself, we have 18 of presence compared to, for example, AWS, which has two Hyderabad and and, Mumbai. So we have 18 just to just for comparison. So, obviously, you have much more options to connect to wherever you want to take your data, but you can keep it with us. Good question. Do Cloudflare mass data at the source node? Okay. So so there are two things here. At the source inbound, yes, we by the time it hits the log, we will mask it. So if your question is SIEM compliance, yes, your data never gets locked logged with PI information. We will mask it. Now there's another thing that if somebody's entering the information in the in the text box would and they enter it, they will see their own response. Right? Let's say a chatbot. Yes. There we won't because that's the personal view. But in the log, it will be masked. So depending upon where you're looking at. Are you suggesting all the Internet facing traffic needs to route by Cloudflare? That's the add ideal situation so that when all your traffic by the way, I mean, this is not abnormal. We we also deal in bot management, DDoS, and all. I'm not covering that topic today. So we are we are provide our DNS and edge and all. So if and if it helps you, 20% of world traffic, which is bigger than Google as well, is flowing through Cloudflare, just to give you confidence on that. So, yes, endpoint for your egress going to Internet, if you're pointing that to Cloudflare, we will give you that benefit. If you do not want, of course, then there could be scenarios where you are deploying Cloudflare solution and testing in some specific region, then you don't want all your apps and all your devices to point to Cloudflare from the start itself. But, eventually, that is what would give you the benefit I just listed. Okay. Does AI gateway to be in DMZ for an air gapped environment? I will need to check about DMZ. Normally, it is something that we work with the Internet, but DMZ, air gapped environment, we would need to put some routing, I suppose, but I don't have a correct answer for that, Manish, in case you can. I'll send my email address pretty soon, I think, on the on the next maybe on the question answer slide. I do have my email there, or maybe I'll type it on the chat. Let's let's do it now. For those who are not or not getting the answer, this is my email address. Do send me your question that I could not answer. I cannot ask your email address, otherwise, everybody will see it. So do send me a question there, and I will answer from there. Right? Gaps, we have. Yes. So the very good question. So everybody's benefit. There is a great question there. Does Cloudflow help identify the gaps? Yes. That is what is called as visibility. Within a week, we will tell you shadow AIs, which is scary. I'm so sorry, but but that is what it is. We will tell you which, apps are not as per compliance, where we are, what is blocked, what is not blocked, where the PI is flowing, which is not as per compliance. Like it or not, we will give you that report. Within a week, of course, we we are able to provide. D b d p act. Okay. So d b d p act says blacklisted countries. Right now, their stance is you cannot transmit your data to blacklisted countries, but there is a problem there. Good question on that, Sharma. For an example, Aadhaar. Aadhaar is not just about one entity. Right? It is connected to tax ID. It it has what is it called? UADI ID, which has different laws. So while DPDP might only be associated with Adhar, you have other things which are linked to it, and that's where a lot of other regulations come along. Safest approach to that is in country. Data sovereignty, in country processing of that data. Simple as that. But, yes, if you say, if you say DBDP, it says it cannot go to blacklisted countries. Simple as that. How do you distinguish between governance? Tooling standpoint okay. The question okay. I need to click on read more. Shadow. Okay. I need to, again, read the question. I'm gonna start in between shadow AI and by using unsanctioned tool and sanctioned, but ungoverned AI is an up to tool without ah, so so that is the policy. So the question is how do you distinguish between tools with from a shadow AI usage, employee using unsanctioned tools and sanctioned, but ungoverned. Right? So we have this is all policy engine. This is where you will come and say, yes. I have allowed for the first time when we are setting. You'll say OpenAI, you you have allowed. Entropic, you have allowed. Gemini, you have not allowed. This you will tell us now because all your intro traffic is going through us. We intercept apps. We have secure web click pay. We have other mechanisms. No agents, by the way. We would know if somebody is using Gemini, and we will tell you. You have an option to block or flag. Flag as in you will just have a red logging versus block. Right? That's the last question. How do we, do runtime guardrails work? Yeah. That is the question that I'm going to answer pretty soon. Okay. Quickly, not just not just logging, not just helping you with compliance. I know the cost was not something that all of us wanted to hear, but then I wanted to cover this pretty fast. There are two main areas where Cloudflare reduces the cost. Number one is token caching, and I have a slide on that. This is a real use case where one of my customer, of course, not publicly allowed to speak. That's why I'm not bringing it. There are 10,000 average, 10,000 requests per day on their chatbot. And once they bought this AI gateway, they were able to reduce about 89% of their, token. How? Because of the caching. Because of the repeated duplicated or so let's say semantic. Earth is round. What is the shape of Earth? So Earth is already something that we could, you know, shape is something we could do as, caching. We could respond from caching, and that's why the speed increases, and there is no token cost there. That is one. And second is GPUs. Look at any hyperscalers. If you want GPU, they ask for capacity guarantee. When you look at reserved instances, for example, they say one year to three years. Unlike those in in CloudFlare, we only have pay as you go model where you only pay when you use the specific GPU. Right? Wanted to show you an example on chat. How does it look like on token? Might sound like a busy slide, but very simple. Any specific request and response on chat have four components, system prompt, context, query, and response. Now first time you send a new query, nothing would happen. If Entropic or OpenAI was charging you $1,007.50, we will charge you the same tokens. The benefit happens from the second time when while system prompt and, context is cached, we already have them. So you will say about 90% here. There there's still 10% that will have to flow even if the query is different. Right? First time you asked, what is the shape of Earth? Second time you asked, what is the shape of Mars? Completely different. Yes. Some some similarity, but not common. We will be able to say about 77% of your tokens, which is humongous in that scenario in that case. And the good part, we work with OpenAI, Anthropic. So so there is a lot that you can play with. Given you have so much of cached data, you could play with the scenarios are a lot where you could think if it is a complex question, we do not want to crash. If it is a simple question, we want to use there's a lot of, area or room to play with. And if there is a same question, I mean, you say, what is two plus two? And second time as well, you ask what is 2.2 plus for an example. Similar query, you will save about 90% out because 10% still have to go through and processing has to happen for. But in real time, I have seen myself. This is my customer where they had average about 10,000 odd request every day, and they ended up saving 87,000 a year from Frontier model usages. Right? Yeah. Quickly, I'm not gonna take so much of time on this, where, we can do routing based on data sensitivity. Route to a specific region player, pair and block mean to say if data is sensitive, route to, let's say, your on prem environment or route to, let's say, somewhere else because you have sensitive information coming from specific team. Now how do we decide on sensitive information? Keyword matching, prompt length, classification of model classification of user who is sending it. You have a CXO whose request should always never land into some sort of, proprietary model. We cover you for that. If query is complex, then send it to, let's say very simple. Sorry. When query is simple, send it to, let's say, open source model, which we you can run it on run it on workers AI as well. DeepSeek versus Anthropic AI. You do not want everything to charge you $100 when it can happen in $34. We all know about 90% cheaper DeepSeek or Mistral or anything else that we want to. That routing automatically happens from from our side, Cloudflare side. We could put budget. We could downgrade a model. We could upgrade a model depending on how you'll so if you look at it, there is lot of this if then simple configuration that you can do. One setup, it is an automated approach where what I have seen personally, 80% of the task become routine. Even if something was complex earlier, the more it is asked, it become routine. The more it become routine. It doesn't need frontier model most expensive model to answer. It can happen it can happen from open source cheaper model, for example, which are free to download, by the way. And complex continue to go for expensive model for such, right, which remains 20%. Now there was one question that I see now. I think there's any other questions. If we already have caching and semantic caching component, will it be ignored? Or, I don't see the full question. Yes. So cache hidden cache, miss concept if you, know about this. If we have it in our cache, we will respond it from our cache. That's the whole point. Because we are caching the response as well on the previous time. Right? So you save on token, you save on time. I have done it myself, like, 100 word with with a PDF document, which, you know, there is similarity caching. Not everything was cached there. We could get it at 77% discounted in terms of token, and it was faster. Previous one took about seven seconds. The second one took under one point five second, if I remember correctly, because of the caching that we had. While the question was not exactly similar, video document was the same. Right? Yeah. So guardrails. We do, let me check the question again. The last one, I I I can't I don't know where my guardrail, slide went. I'm so sorry for that. But then how do runtime guardrails work? Right? This is a this is the question that came first off. We'll what is runtime? Now runtime is where we are in between your origin and wherever destination it needs to go. Right? We look at, for an example, bad traffic, which I'm not covering today, by the way, bot traffic, which has surpassed surpassed human traffic, by the way, just for clarity for everyone. We look at that. We look at Bola, if you understand that. We look at JWT, for an example. Now those while, other organizations like AWS call it as guardrail, we call it as WAF, web app firewall, or we call it as our own DDoS protection or, you know and we protect that way. I hope that answered your question. Benefit of time, I'm not going too deep because I'm left with only two minutes there. Quickly switching to my last slide. Right? So, if you start with CloudFlare, you get visibility within one to two weeks. You get who is using what, where, why, those kind of simple questions immediately. You get control or rather you have to, of course, configure control and you will have more control on PII, more control on routing, more control on cost, that you have to configure. Visibility comes as default. Then you can apply enforcement. And the third one, logging, of course, it is optional. You may or may not want to do logging with us. If you do, there are separate benefit. If you don't do, if you have your own chosen field, you can go ahead with that. With that, I want to bring my, slide which talks about, of course, waiting for not having visibility. And visibility, more or less, you can test it for free, and you would know the benefit within one to two weeks. I think no harm, no barrier. Of course, after you want to have advanced topics address, it might cost you, but not before. Questions, please reach out to me. My my email address is also mentioned there. I know I ran out of time. It's almost top of the hour. But I still welcome anybody who wants to have q and a. Otherwise, we are almost on time. Vinay, what do you suggest? Hey. Hey, Shadi. Thank you. I think I have extended the session for five more minutes. And I see there is one more question left from Andy Wong. He's asking, how does CloudPass AI gateway help institutions maintain an audit trail for model usage, which includes inputs, outputs, and decisions for regulatory examination purposes, not just security monitoring. That's a good question, Andy. So, for those yeah. I mean, please, feel free to post your questions on q and a if you have. But I'll address how does, AI gateway answers. Right? There is there's a big question and a little big answer, if you don't mind. So AI gateway, first off, gives you visibility, as I said. Secure web gateway, then interpreting your APIs, interfaces application interfaces, it collates all that information. Right? So and then it starts to filter out in the buckets on showing you which model, where, what what was used. We have the data. All we got is dashboards where we start showing you now. So that is part one. Now part two is how does it help with regulations? If you have been with any hyperscalers, this is something that is shared responsibility. I don't want to say we also have the shared responsibility concept because that might sound like a I'm using somebody else's system. We provide you with SIEM consistent logs, which is what predominantly the first thing that, are required by regulators, for example. Now you need to build out based on because I cannot cover all the countries in this session. You then need to build out reports. You then need to build out your own mechanisms or or responses. Once you have evidence of the data in a format acceptable by regulator, you cannot just simply, you know, take all the logs and pump dump it onto the regulator. You need to then massage it as we say, create a report insightful report and tell them this is evidence, reconstruct it reconstruct the incident that has happened and tell them this is out. So so if if I answered your question, let me know. If not, if you have a follow-up, repeating again, very simple. We have visibility because of all your traffic is going through us. We collect the data. Thanks for that. We collate that information, and then we put it in SIM format, which is where I was talking about masking off the information and such, which is at the back of what regulators ask. Now for everybody's benefit, this SIM log does not need to be in Cloud Fair. You do not have to pay for it. If you like to have it on s three and then you want to use something else, you might guess. Never mind. Yeah? But then, yeah, we give you visibility there, and we give you a lot of intelligence up from the very beginning. Right? Okay. Thanks for that, sir, Shakti. There's one. more question from Ramesh who is asking, does this need to be installed no. Sorry. I'm sorry. Does this need to install any agent on endpoints? Good question. Okay. No no agents. We do not have any agents, in most of our solutions, not just, AI gateway. We have about what now, close to about 50 products or services, give or take. This happens via API. So all you got to do I will be very clear. It's not like it it it is a magic. You divert, and that was one of the question, all your traffic through us. So I can bring that slide just for the sake of clarity. Yeah. So you configure an endpoint URL that all this traffic, either your employees or your applications or your APIs, everything is diverted to Cloudflare. Then and only then we will be able to cover it. Then we would know shadow API. Then we would be able to do give you visibility in that. No agent as such. It is egress only. We do not have an agent which will collect the information, pump the log like many other telemetry solutions out there. I hope that answers the question. Alright. I think with that, we are at the end of our session. All the questions have been answered. Thank you so much, Shakti, for taking the time out. Also, thank you for all the attendees for joining today's session and interacting with us live. Hope we found today's session relevant and valuable, and we look forward to seeing you all at an upcoming webinars. Yeah. Have a good day. Thanks, Vinay, and thanks, everyone. Have a great day ahead. I appreciate your time. Do remember my email address if you need to reach out to me for something. Otherwise, have a great successful day. Thanks a lot. Are we in the backstage?