AI Observability: The Missing Runtime Control Layer for API-Driven Systems | apidays India 2026

can start. Um, great to see you full room after after uh lunch. It's usually harder. I have to keep the people awake or try to. Um, so yes, thank you very much for coming here. My name is Is Gross. I'm the CEO of Amunai from New York and I'm very happy to be here in India for this event and also for all my meetings. So I'm very happy to be part of this uh journey. I also was in uh several API days around the world in New York and other places. So a lot of experience and I see this a lot of people coming here in afternoon. It's makes me happy. Oh, so thank you very much. I believe all of you are interested in API security. That's why you all here. And let me uh give you some of the information like we can uh talk about today. is about it's about non-human identities and what's going to be now with all the API security and that [clears throat] so so my name is I said my name is Gross I'm the CEO of Amunai a New York based company uh we work with a very highly regulated sectors like banks governments and others in order to make sure the their all the environment is secured And also we have a lot of experience working on the compliance regulation around the world to make sure that those companies are complied to that. We're also part of these compliance. We part of these writing this compliance for around the world. For example, the FDX in United States, the financial data exchange, we're part of it to make sure that all the banks are in the same format of security. So we also sees the governance side of that. So what we're going to talk about today is going to have uh the shift that's happening now some case three sweet case studies uh how we see the control model and the road map to actually be more secured. So first of all uh what we see today is um the actually the security now is not AI apps using it's more about those agents that have a lot of uh flexibility and they have a lot of responsibility. So also the AI agents have a lot of uh role today in the today in every organization and I can see example from in UAE they decided that AI will take decisions okay it will be like employees so AI will be part of the team and AI can actually take decisions even uh was a use case this week that the AI agent fired as a fired an employee a real employee a agent fired employee happened this week it was published this week at least Um it's an interesting story if you want me afterwards. So those non-human identities we're going to talk about what we saw from black hat. So I came from black hat in the United States. I was seen some examples there like everyone was talking about this AI security and how AI security will come to organizations. We're going to talk about some of the permissions and all kinds of AI get and sometimes have over permissions. We're also going to talk about identities. what used to be one identity tomorrow can change and this has to be very dynamic. So some of the points we see here so today the ratio between human and nonhuman is 1 to 45. This means people connecting systems 145 is human. uh all the rest are nonhuman and some web systems app system it's 1 to 144. It's dramatically a big change in a market now. It's not any anymore you're not anymore securing the people you have to secure the nonhuman identities. So you when you look about those agents um those that can do have a you have to understand the reasons about the task why is they doing it how how's they doing it and which tools or request credentials for each one of them. Um some of them they can be uh they have sub agents so you have to manage like AI agents in the sub agents as well and and they can have access to multiple systems. So the AI is actually much different than a human. You know, we all know that. But just from perspective of of looking how to secure it, you have to see it as five different aspects that are uh dramatically different from human people logging into the system. So of course the ability that they can do the mistake its mistakes can be in scale because a person can do number of requests but when the AI agent does a mistake it can do it so many times and it can cause big damage as well. Um I want to give the first use case is that the audit was running on the system and they identified they identify those AI agents and we identified it's the AI agents of course those AI agents were approved and and and seemed to be legitimate but then actually no one was looking what they actually did. So let's say I I have a a agent that he's supposed to do uh register a new user account to the bank or something like that. So the agent has ability to do it but who knows what he's doing besides it or he's doing other things. So we know what he's supposed to do but we don't have the full visibility on what he's actually doing. So we can tell him what to do but if he does more than what he's supposed to do that's could be a big risk for organization. And when we're looking about this, so this organization had, you know, the policies, a logs, IM groups, API gateway, access rules, a par reviews, they had all the things in place to be secured. So we as a problem is they had everything but they couldn't understand one thing. What they were missing is that what was behind the intent of those requests like what's actually happening behind the scents and they didn't understand if he was doing what he was supposed to do besides that he was allowed to do but what else he was doing. So this was something very interesting. We saw that case from compliance again. They said this is a API allowed. Is this AI a is this AI agent could do things? He's authenticate. We know who he is. You know what he's supposed to do? And what so what we we think we they need to add is a point of the so for example make sure to put guidelines around it. You cannot just say I trust it and let him do whatever you want. So you have to give actually audit the specific behavior and usage of that AI agent and that's a very interesting point because think about today it's it could be like actually chain here that happens here look what we can see here. So it could be a human request and then a uh a um he agent identity goes to do it then he has the MCP server API action and goes to the data. So some of the some of the risk here we see here for organizations that they're not really aware of the end user behind the asking the request okay because it's behind all these masks. Think about this and you have you have so many gateways like or gateways who can say oh say entities that use the same identity. So for example I'm trying to get apply for a credit score and to a bank and then I go to one of the system and it applies on my name to others and then also the side who accepts this information realize that that's a real human behind it. Okay. So and they start all talking together but you know you don't have all the 100% that this is the user behind it and really user really user that. So this is very interesting to see that that we all build everything is built on trust and it's too much trust you know when someone can actually manipulate it and that's that's where the big risk happens for organizations. So we have here um the the flow and you want to make sure that this is all within the behavior of actually was permitted and acknowledged and all these. So you have to not only as to summary about this use case. Yeah. Not only about knowing who's a user but actually making sure that he's not going out of his boundaries what's not supposed to do. So that's about this use case. We can have another use case of um over permissions. Very interesting you know like uh the AI agents takes more responsibility. He becomes a very loyal to the organization and he starts to take more actions that he's not supposed to. Very interesting uh uh use case. Um so we can see here like the where where um agent comes in it can read update and all kinds trigger all kinds of work to be done but we want to make sure that he's doing exactly what he's supposed to do. So for example that he know he's doing it specific as his identity and what's the purpose and scope he's doing and if there's any sensitive more information he's asking now we want to make sure that it it's being under control and again also we always have to validate the policies. So think about this is you have like we call it like a new employee in your organization that he's doing all kinds of things. So this AI agent and we're going to get back to he's like a new employee. You have an employee. He has you have to make you have to monitor him. You can't just drop and go. And it's so cool because we all can develop these agents so fast and we have to make sure that they always be under control to do exactly what they're supposed to do. And a third use case is where the AI goes outside of the expectations. So we we the three use cases are in three levels. One is to make sure that even the same AI agent that does on the API calls he does what he's supposed to do. Yeah. You have to still maintaining look on it all the time. The second one talks about if they get, you know, a more responsibility and taking extra um extra care. And the third one is they're just completely going out of the task. You know, they they're defining themselves their own tasks and ability to do a different thing. So, so you have to make sure that they all playing in the same ground. You cannot let them go out and and do all the things. So um so from risk perspective once a AI agent and we all know all the people here in the room are very smart and so AI agents have capabilities and you want to make sure that uh not getting them over permissions to do all kinds of things because they're going to find the loopholes to actually maybe create more capabilities and to get more access and to change data and etc etc. So you want to make sure that all of this is also under control. They don't leave the the area. So these are the three types of use cases we see and uh and uh now is also very very focused now on all the AI security now today in the market. So when everyone talks about AI security they're not talking about securing open AI okay the modules itself everyone's talking about how you secure organizations that's most of the let's say 95% of the AI security companies are talking about how to secure organizations not to secure the models so so when we talk about all those AI agency they all go on APIs so this is one thing is clear so everything is API this is a good standout you have to start so you have to once you have to you have to you see those API You want to make sure that those AI agents are actually exactly where they are, what they're supposed to do, and they're not going outside of the responsibilities. So what we see here, so today we see the logs, we see the endpoint method, the status code, identity log, the token principle times, and of course gate, the gateway logs, rate, IP route. That's what we see today. But you want to get additional control is um like which agent took this action. Okay. So we can put the task per agent. Okay. Who and which users behind it. Okay. On behalf of which user is is processing this data. Okay. And when one user different users um which business intent like what's business behavior is trying to do what is he actually trying to do? All this you need to. So this was in the past. So the past was the left side and the right side is what we have to go now. This is where all the organization are focusing now. A which data is going collected by AI agent. It's sensitive not sensitive. Um if what to build a baseline of normal behavior and then to see if this changes from the behavioral and of course um do we allow this action still yet now? Maybe it was allowed yesterday. So you always have to monitor this. It's not maybe if the agent who had the permission yesterday but today is not allowed, he shouldn't be able to do that. And when you look about a observability uh we want to find like this the five we call the five points. So first of all you have to map you have to discover all your uh APIs running there with AI tools around them. So you want to first discover everything once you discover like for example which u agents APIs if it could be third party because a lot of third parties are using the a AI agent. So you connect your system to a AI agent. This is actually AI behind it. Someone's else. And the first thing is to discover everything. You have to have the the most important today is AI governance. It's the ESO 420001 and people looking into the EU act. This means new regulation for AI. So you want to have first vis visibility of everything. That's discovery. Then binding means okay let's try to give a all those request to give them some um I would say give them some some information about them for example so I want to say about this a a identity what is permissions what you supposed to do and what's the purpose of it after that we correlate it with uh the make sure that we have this map of this API we have it where exactly it's supposed to be that we don't have any mistake that's the whole flow is part of it and then we need to decide. So every organization has to take his own decisions like what would be the governance of those um APIs like what exactly he wants to do and to allow can be different policies for different organizations that's why you don't want to have a specific uh rule for everyone cuz maybe for this organization is allowed for that it depends on territories and sectors it could be if it's a cloud environment on environment this and then after you decide and you build the rules then you have to actually enforce so this is the five things I can tell you everyone's still stuck at number one okay to actually fully discover that's where everyone's stuck yet. Once they know everyone where's the discovery and that's why the regulation is pushing very very fast that everyone have this ability to discover then you can actually take the next steps and of course understand what's going on and take control of it and actually decide and enforce at the end. So um let's talk about the six controls for non-human identities uh today. So one thing is about unique agent identity. You need to have each one has to have its own identity. You can't have like a one and he has diff different capabilities. He has to have each one has to have its own to put them to put them in in actually in order. So you have to have each one has identity. Let's say we have employee number. So we have agent number. Uh we can go to the allow allow list. Okay, we build allow list say okay which tool can do what which AI can do what and you have to after you have the number you put for it what's this task was we're supposed to do then you say okay does it have sensitive data it doesn't have sensitive data does it have some information about the you want to you want to make sure that we want to move we move those sensitive data we allow it and that's that part um the fourth one is short live credentials so AI agents cannot live forever Okay, they have to have be time scoped so we know it won't be running in the system forever. It has to be controlled maybe you know renewed or but has to have a deadline for that and we so we don't find after two three years that the AI was running there um limitations so we want to build limitations uh for each one of them um for those API calls like what's what's allowed to do or not to do and human approval thresholds this is funny that we're going back to people right now we're going back to the human people we talked about a agents but someone has to take responsibility on this and say what yes what no it's not a trivial that that they will do everyone someone has to approve it because then they were going to come ask one day and you can say we were breached on this API yeah but who took control who allowed this has to be of course part of it um going next the audit okay so from static compliance we have to move to runtime operating evidence in the past was mostly about access metrics policy documents ment quarterly review. This is what we did in the past like very uh very uh systematic architecture diagram sample logo explorer that was the old world. Now today we have to go to the runtime operation because this is very live you understand if you do it one once in a quarter it's not relevant anymore things can change hourly. So you want to have something that's all the time going we call it the runtime runtime operating evidence. So of course for we talked about the trial of those identities um we're supposed to do the sensitive policy enforcement that's always the five things we talked about in the beginning. So that's the shift the move today. So all the organization now are taking these as action items moving from the static to the runtime security. So what do we suggest organizations do? So we just suggest a 30 60 90 days. Um it's pretty recommended for organizations today that have AI agents and the first thing is to build the inventory. So take 30 days build inventory find all those AI agents BS API keys map them which data goes on them and and identity or identify also the business owners and purpose of those. Then you have 60 days uh metric is to actually uh create the rules for each one of them and also actually remove standing privileges. Okay. To make sure that we have control of this and to see and also to map them of the high risk. So we can say this is high risk you want to monitor more. And of course the 90 days is to enforce all the policies to make sure that all the policies that what's allowed what's not allowed is inside organization. So and if you're going back to like the five questions to take back and you come back and work on Monday. So five questions first you want to ask your organization is uh which AI agents and automations can uh can call production APIs today. Uh which of those agents uh use shared longived or overcoped uh credentials? Uh can we map AI uh can we map um a agent actions to human people? If not, which API actions are actually here uh for financials privilege or sensitive so you can actually put the scoring risk on that and where are we going to enforce so these are things every organization needs to ask itself these five questions in order to be u more secured and I think I'm coming to my end of my um presentation um so it's not about approving the prompts not approving the prompts It's more about governing non-human identities today. So this is where the shift world and if you can take one thing from this uh this presentation is that being part of this transition is very important to notive because it's a very special I think we all here now in a unique generation now that it's something happening in life and we can maybe take it to our kids and talk tell them about this. So a [clears throat] this is a very good timing. We have to make just make sure that's under control and thank you very much and looking forward for any questions. Thanks. [clears throat] Yes. >> Hi. Um thanks for uh these insights. The questions that you showed they are very interesting but my question uh about that is which are the key roles who needs to discuss those questions together because in most organizations um discovering the system itself is different picture for different departments different roles. So do you have any suggestions on which are the key roles we need to discuss this together? >> You getting to a interesting point you know the battle between uh the developers and security. [laughter] >> Yeah. So, so developers say we have everything under control, we own everything, we know everything, but security want to make sure that that's okay. So, usually this will be under security because security matters. They want to make sure that the that the the API developer teams are doing the right things. Okay, could be everything's okay, but they have to get control and visibility. So the security team it's kind probably under the appsec team that they want to make sure that all of this are complied and then they report to the security uh CISO GM and they can also afterwards share it with the API developers say okay our recommendations would be to to scope those APIs to put some uh some uh I would say some gateways some uh barriers for that and etc. So that's how it works. >> Thank you. Yes, >> start first. He has a he has a microphone already, but yeah, >> I just stole the microphone first. >> So, I just had a quick question. Um, do you see the governance engine or the governance policy that you just showed being deterministic in nature especially as the AI actions grow or AI agent multi-tool actions grow? Have you seen that kind of implementation anywhere? AIA governance >> that yes a governance >> AIA governance you ask for specific for Asia right >> sorry >> you're asking for Asia right yeah okay so US everyone is putting out a governance it's part of the IO4201 so it's a comp everyone has to comply with it >> in Europe they putting enforcement it's called the EUA act and everyone has to have control on the AI so they all have AI governance now mandatory it's not because they're being in risk because it's mandatory this is where very important is here I can tell you in in Asia it's less we are still focused here in this market the API security and the basic but this is going to come the next step so now when we work in customers here in in in the region we already gives them already the AI governance as they will need this the day after but it's just it's just not yet in this market where US and Europe is already in AI governance yeah >> is it not just human also software based governance like are there platforms being developed to actually govern the agent actions things like replaying uh the agent's action and you know full auditability and things like that are are those being implemented in Europe in in a software platform >> yeah yeah it's pretty it's pretty uh common today the organization are deploying um AI governance as a software not as a service we also have it as a software it's a software you put it in organization it just it does everything for you the whole reporting and everything the first part of the discovery you know That's what the every what is missing. It's called AI governance. That's where you start with and then the organization has to take the actions and decision. It comes the next steps afterwards.