Well, welcome to another Akeyless webinar. And this one, I'm actually very excited about because we get to talk about securing AI agents. We're gonna choose Claude for this webinar, but, obviously, it can be with any other AI agent that we have. And, I'm joined with Barak. You probably already met Barak. Barak, you probably did some introductions, and we do have an exciting demo for you at, at the very end, so we're gonna get that started soon. So let's go ahead and look at the actual, issue or challenge that we we have here, and that is that agents are getting real access. They're getting access to your production systems. Believe it or not, these AI agents, whether built on Claude, OpenAI, Gemini, or any other model, they're increasingly connecting to your databases, cloud services, SaaS applications, through MCP or even through the vendor's CLI binary. And too often, that access relies on long lived credentials that are sitting whether in MCP config files, in environment variables, or inside the agent's memory itself. But the key here is that these are standing targets, and any attacker who reaches that agent environment is able to to access those credentials and, even perhaps access your systems. Alright. Now research consistently has shown that compromised credentials are involved in the majority of successful security breaches that are out there. Even more concerning, most identity related breaches now involve nonhuman identities rather than human users. And, maybe you can see the the the last three here on the list, they involved AI, and the rise of AI introduces a new dimension to this problem. AI agents are frequently requiring access to multiple enterprise systems that rely on static credentials. Worse, an AI agent with excessive permissions may take actions that are technically authorized, but they're operationally dangerous. You're gonna hear us talk about intent quite often, and this is why identity security must evolve beyond just storing secrets in a vault. Organizations need to focus on runtime controls as well, verify the intent, issue temporary credentials that expire after a certain time, enforce lease privilege, monitor actions as they occur, and eliminate any standing access wherever possible. Where credentials exist and what can go wrong on the diagram that you see on the screen for you here, the we start off with the the user, and, basically, this is where the AI agent execution loop and where the key security gaps may appear. So there are different places where this can appear. You can see here at the bottom the legend credentials present and attack surface without control. But we start off with the user who has a request and, but their identity can be lost when the agent acts through a shared shared service account. And then up here at the top, the LLM itself can be influenced by prompt injection from emails, documents, tickets, or other untrusted content. And then the agent itself may also hold that those long lived credentials, API keys, tokens that we talked about, increasing the blast radius if it is actually compromised. And down down here at the memory and rag layer, this can introduce poison data or expose sensitive information to the model itself. And, with traditional RBAC, you know, RBAC controls what an identity can access, but not whether a specific action is actually appropriate. And then finally, the tools, APIs, MCP servers, they execute actions across enterprise systems without a central enforcement point so that audit logs may show only the service identity instead of the user and the the the user prompt and intent behind the actual action itself. So when we talk about, maturity in terms of securing your AI agents, we find them in two stages. Normally, the stage one here, secretless AI, and then stage two, runtime authority. So the first piece here as you adopt AI security, or AI agent security is we want to connect without passwords. We wanna eliminate API keys from prompts and use short lived credentials. Right? And then when we talk about runtime authority, you can see here some of these, points, execute production actions, validate user intent, enforce runtime policy, mask sensitive responses, and stop unsafe actions mid session. And, you'll see Barak when he, shares the demo, you'll see that this goes both ways. On the request side, you want someone in the middle to broker and to see what's going on and allow depending on some policies what goes through. And on the response side, you wanna be able to also, check the responses, mask any sensitive information, and so on. So digging a little bit deeper into stage one, secret list AI, we have where we start off with the agent request access, and then Akeyless brokers the credentials, and then the credentials expire after a certain amount of time. And the gateway really, the Akeyless gateway is the one that creates those short lived least privileged credentials, they're created only when the agent needs them and they expire after use. And your LLM never holds a real credential. It just asks and Akeyless brokers the access. And when it comes to runtime authority and enforcement, that stage two and Akeyless agentic runtime authority that's brokered through the gateway, it works alongside your MCP connected AI agent, again, whether Claude, OpenAI, Gemini, and it brokers access to your enterprise systems, not just at login, but at the moment of action as well. So it's ongoing. Right? And it's not just the first prompt or when the agent authenticates, it's ongoing. And, the the, Akiles Agenetic Runtime Authority, it allows intent aware authentication, dynamic access, real time checks to protect your production systems. It's gateway controlled, zero credential exposure, and gives universal governance. Anything you wanna add here, Barag? No. It looks exactly as I start started to describe before you joined. We talked about the secret list part. And now Yeah. Definitely, this is the the next level where you're starting to also enforce the policies. Right. Right. Yeah. And, really quickly how it works, you'll see it in action. But just quickly on this diagram here, you see that user prompts an AI agent for a data request. So you see that goes through. And then the request routes through the customer deployed Akeyless gateway. So you deploy your gateway in your own infrastructure, and what happens is the gateway evaluates the request against a set of policies that you put in place and then if approved a just in time credential is issued from, that single action and then the activity is logged and the credentials expire immediately after use, so there's no long lived credentials. And then on the return path, the gateway once again evaluates that response before finally returns it to the agent, and the agent outputs the response to the user. And Akeyless has this connector to where it allows you to configure to configure it through MCP, and it brokers access to all your, systems, the databases, cloud providers, SaaS platforms, which helps you replace a sprawl for your you know, per system MCP server, it's done once. So what's the outcome? We get, three outcomes here, one for Sequel's AI and and two for runtime authority. Sequel's AI, you get zero standing credentials, no more credentials sitting in config files, environment variables, or agent memory. And, runtime authority, we get two benefits. The first is intent aware enforcement. Every requested action and response is evaluated through policies, and then we also get a very nice forensics traceability, capability in a dashboard in Akeel's where you can actually see every action log tied back to the originating prompt. So with that, we have a live demo. And, Barak, I'll pass it back to you. I don't know where you stopped, so maybe we can pick up, where you stopped. That would be great. Cool. Thank you. So I'll take it directly from here. We just ended talk about the secret list, and now, as mentioned, talk started talk about the runtime authority where we also control not only the the itself, the secret less part, but also the part that we control about the policy, what is the in or the out policy that we want to define. In this demo, I'm going to demonstrate an AWS access. So I've created a dynamic secret, just in time secret to my AWS account. This will be with excessive permissions intent. I got myself some EC two full access. And this is in order to show you that, basically, upon that permission, the authentic runtime authority rules is above that. It's blocking us even if I'll try to manipulate it and try to get some more thing that I can do. So as I mentioned, we have the input and output rules. Because it takes some time against AWS, I will start running it. So first of all, just to explain, we have our own extension for Akeyless' agentic runtime authority. It is configured with my OIDC, so it has my own identity. There are no secrets inside, Claude itself. Of course, you can use this different authentication method supported by Akeyless if it's for human or nonhuman. And then we have the option to run our command. It will start. It will try to fetch items from Akeyless. Probably, will ask for permission first, and then we'll take it further. And then I will show you some more information from Akeyless' side during this time. So now it tries to connect. Okay. This is my authentication. Authentication succeeded. Now it's going to Akeyless. And now let's give it some option to run. It will try to list EC two instances against the secret that it found slash a r a AWS dev exactly as I have shown. Here, this is the item itself. And now we should start seeing some prompts. And by the way, and very nice, I found out today, you can also have a debug option by seeing inside Claw directly all the operation being done. So once we will have some more information, we can see it here. But also okay. It's running, so we can see. Okay. So I've asked to see all the e c two machines including the internal IPs, and I got all the relevant information. The next part will be some restart to one of my machines. So I gave it some information. One of the machine is stuck. I want it to be restarted, and now it will try. And first of all, we'll check the policies. So if the policy will allow it, it can continue. Okay. It looks fine. Try to stops now. And if we go into Akeyless during this time, we can see that we have a monitor that shows us all the operations being done, if it's blocked, if it's not blocked, what was happening during this time, what was the prompt that it was running, and what what was the output. So we have a forensic traceability all the time. That is incredible. I love that traceability. Yeah. Indeed. It's amazing. But what I found that, as I said, I found that this plug in also allows me to see those operations directly from here. So I'm not me as a as a developer, I have the ability to understand what's the problem, what I was asking wrong, or what was running, and then I understand all the relevant information according. Okay? So you see I got rejected by the AI policy. It will start trying that again. Something didn't work for him according to the policy, but it will keep trying according to the permission level he has. So during this time, let's go back to Akeyless. Again, the item itself, as I mentioned, has created a just in time access. I can go to the temporary credentials and see that these credentials has been created by me. Okay? I was the user coming from Claude and asking for this permission. So everything is traced and yeah? The attribution is there to you. Exact exactly. So I can see all the time who is the user, who what he was asking, what is the trace of any operation he was done. Okay. So now the machine managed to stop. Now it will go and try to start it. The next prompt will be trying to stop one of my production servers Where, definitely, it's something I don't want my agent to be able to do. And by the way, the delays that we see here, this is an artifact of AWS. It takes time. Exactly. It's against AWS, so it takes time to validate the operation against AWS. It's also waiting for the server to stop and then start again, always monitoring it. Yeah. So you see it. That that is why it takes more time. Okay. Now it's waiting for the running state for this specific machine, which is fine. And the next one, as mentioned, I said, I'm telling me the production is down. I wanted to stop one of the servers immediately and do a restart. This should, of course, getting blocked. So let's see once this one is done that we indeed get this result. It's interesting because as as we're looking at this, I'm thinking in the back of my head of infrastructure as code. It looks like with the ability to have security taken care of, I can I don't need to run Terraform? I mean, this is this is probably a bit of a do it for you. Exactly. Like, it's you just need Claude, and then the agent takes care. You just tell it what you want. And since security is taken care of, that's very interesting. Okay. So I've run now. Server is running, restarted successfully. Now I'm trying to stop the production server, but production is stressful. It's testing again the policies. It wants to make sure that it can do some operation against it. It should fail, of course. Okay? And it says, I'm not going to stop this one. Okay? You can definitely talk with him. And the next prompt will be, or the last one here, will be to try and manipulate it. I'm going to tell it that, I have an approval for my CSO to to do this operation, and I wanted to do it anyway. Like, remember that I have EC two full admin. So trying again, trying to ignore the rules, it won't let me. Okay? The policy is blocking it. And if I'll go back again into Akeyless' agentic runtime authority, I can see the operations. I can see what was running. I can see what was blocked and what's the reason. Some of them will take a bit more time to appear here. But as mentioned, I can also see the things directly from here. So I was trying to run, and you can see credential not exposed, but I'm not able to do the operation because the policy is blocking me. This is mainly one of the things. Now think about any other option, that might be interesting in your mind. For example, if you have a database and you have some sensitive information in database and you don't want him to expose, you just need to mention it. You just need to tell it, do not expose this relevant or this sensitive information, and it will be blocked, and you will not be able to to show it. And you do that in the policy, in the output of the policy. Exactly. And it's a text free. You can just write it, and we added some more polishing option to the prompt so it or we'll try to understand you better. And we we're adding some more options and some more improvements all the time there. Very cool. Are you able to show us the policies again in the queue? Of course. Let's go to the item itself. So it was the AWS item. Okay. So I gave it again. I wrote some input policies, some rejection, don't allow destructive or provision operations. And then if there is some sensitive information such as account ID and ARN, I also added that it will not expose it. And some production environment, as you saw, it's it will not be able to access it or do any operation to harm it. Very cool. Indeed. Yeah. Folks, any questions at all? We have a few minutes left. Is there anything else you're showing us, Barak, or is that Nope. I think this is it. Perfect. Any questions at all, folks? You can use the q and a. But, yeah, this is incredible. And this, this feature is quite recent, I I would imagine, right, that Nikhilos is working on or just released? Yeah. It has been released. It's a GA now. It has been released, but it is quite new in few months back. Wow. Very cool. I think that, like we said, this this solves a real problem where a lot of organizations are very apprehensive in terms of allowing their their engineers to access real production systems through AI agents, whether, again, Claude, OpenAI, or Gemini, or what have you. But but with that in place but but the thing is it's so nice and so easy. And if you're playing around in development environments, honestly, like, I don't I don't read logs anymore. I don't know, Barak. If you read any logs anymore, I just push it through. Yeah. Exactly. Look through the logs. Same thing with what you just showed in that demo. Like, hey. Show me the easy to ins I could definitely go into the AWS console and look through and try to do all that. But why? I'm I'm gonna have a bunch of AI agents running in different terminals, and I'm just kind of orchestrating and and directing asking it to do things for me. So if we saw if we can solve this, the security issue behind it, then then why not? Right? Yep. As mentioned before you joined, the main problem also for our competitors is mainly on the place to store those secrets. They also need the secret, and now they need to store it. And now they're thinking, okay. I will do the operation for you like Akhil is doing. But where I'm going to store the secret dynamic, rotated, whatever I'm going to use. Yeah. I mean, Akeyless over the years has built a platform that I don't know if they knew at the time, but it just being built and, staged to a point where it can handle AI workloads in a way that yeah. You built the infrastructure underneath. Like you said, you have a place to store these secrets. You have the ability to create dynamic secrets on demand just in time, and then they're they're gone when they're not needed. So very, very interesting. Very, very cool. Yeah. Indeed. Alright. Any any questions at all? If not, I think we can we can close here. This is such an exciting topic, a really good demo, Barak. Appreciate it. Thank you very much. I'm really excited to see this because I haven't seen oh, well, there's a one question. Hold on. Hold on. Before we close, where'd the questions go? Okay. I don't know if you see them, Barak. I can read them out loud. Do you support Microsoft Power Automate and or Copilot? Copilot is something that we, about to release. It's something on the on the road map. The first one, I don't think we have at the moment. Okay. The next one, just to clarify, instead of infrastructure, instead of infrastructure credentials to sit within the code base slash cloud project, They sit within Akeyless, and Akeyless issues short tokens to the agents, and then it authenticates to the infrastructure. Does this work with on prem platforms too and not just cloud? Yeah. Definitely. It's working with all dynamic secrets of databases, but also we have added lately an option to configure. If you saw, MCP settings for other systems, most of the time, you have the credentials in it. So we have now released some, static secret that you can put your MCP settings instead outside. You can put it in Akeyless and then call it from your, cloud or whatever tool that you're using. And this will do the operation with the MCP of another vendor without, again, exposing the credentials at any time. And, again, it speaks back to how Akeyless has has built this not just for cloud, but on prem services as well. Well, one more question, here from Rashid asking, what about custom AI agents? Yep. This is definitely what I just answered. If the other, vendor is working with an MCP that you can define manually, so you just open an item in Akeyless, put all the relevant information in it, and that's it. By the way, for the first question about other other, plugins or other connectors, it's basically a JSON. So if it's supporting already a JSON in there, it can be running out of the box. You don't need to do anything. Got it. Got it. Good questions, folks. Any any other questions here? Alright. Alright. Well, I think, we can close here. And, once again, Barag, thank you so much. A very good Yeah. It was a pleasure as always. Yeah. Exciting to see the advancement in in what Akeyless is doing and, solving these, difficult problems. So thank you, everybody. I appreciate you, and we will see you in another webinar. Take care, everyone. Thank you. Bye bye. Bye.