Alright. Well, hi, and welcome everyone to another Akeyless webinar. And today, we're going to be talking about how Akeyless brings AI agent identity security to federal agencies. Joining me on the call today, we have Matt Connor, is the field CSO at Second Front, and Ira Shackelford, who is the senior solution architect at Akeyless. And I'm Sam Gabriel, your moderator today. I am a content creator and educator in the platform engineering space. So, Matt, could you please, briefly introduce yourself? Sure. Thanks for having me. Matt Connors, field CISO here at Second Front. Been here a few years myself. My background is extensively in DOD and intelligence community where I previously held leadership roles, as the CISO of the National Geospatial Intelligence Agency, ending my career in government as the CISO of the intelligence community, before striking out for industry. So thanks for having me. Thanks, Matt. Appreciate it. And, Ira, if you don't mind introducing yourself. Thanks, Sam. Yes. Ira Shackleford. I'm a senior solution architect here at Akeyless. I joined earlier this year after spending time in a a couple of other vendors in the space, in the identity space. So I have a a long running background in this. And, before that, I was a practitioner on the other side of the table doing the work. I work here with our enterprise and public sector customers and prospects, to help them, you know, secure, secret certificates just in time access. And now as we're seeing, of course, trends towards more and more agentic activity, keeping those credentials safe and out of agents' hands. Thanks, Ira. Appreciate it. And for folks on the call, please feel free to post your questions in the q and a section. Just so you don't forget, as we go along, we will have q and a at the very end. Alright. Let's go ahead and get started. So recently, Akeyless announced FedRAMP certification class d. So over to you, Matt. Could you please explain a little bit more what that means for our audience who may not be familiar with this classification? That would be great. Sure. Well, just to baseline everybody, if you're if you're interested in FedRAMP at all, it's a fifteen year old GSA mandated program to ensure that cloud security and cloud service offerings, are secure, resilient, and available for reuse across the FedSIV environment. Achieving class d is, an incredible accomplishment for Keyless, built on the game warden platform. Class d certification is what we used to think of as FedRAMP high. It is the high watermark for the FedRAMP authorization, and it reflects the satisfaction of about 415 controls, which you can read as requirements to show that the, application is resilient, hardened, secure, and fit for purpose. So, class d certification, is is the target, the brass ring that most vendors and partners are looking to achieve. You see, on our content there, there's about a 100 or so CSOs have achieved that FedRAMP class d at this point. It is a barrier to entry for those who don't have it to the federal civilian executive branch agencies, but it is also once achieved the the fast ramp to quick pivots across the community. So from a DOD perspective, DOD builds on the FedRAMP high baseline built on the same NIST 853 control set. It's quite an accomplishment to reach class d, so props to Akeyless for the... For being there. And for everybody else who's interested, it it is the the unlock for these markets. Great. Great. Thanks, Matt. So just to add, I'd like to learn a little bit more on what Dow impact level five is and how that ties to FedRAMP class d. If you can touch on that a little bit, that would be great. Sure. I mean, it's natural to wonder sort of the relationship between FedRAMP and the Department of War and their authorization framework works. Department of War follows an impact level model. This is defined by the the DSSA's SRG, the security requirements guide, and it it correlates to information impact. In... Information at, unclassified level without any of the restrictions is impact level two. Information that is controlled unclassified information is at impact level four. And data that is both controlled and unclassified information and includes national security systems information is at impact level five. That's the highest standard in the unclassified DOW world. And so there is a correlation to that FedRAMP high being the highest sensitivity level for the FedRAMP protections, along with the DOW impact level five. And in fact, the control sets, basically overlay with a small uplift for defense specific controls. So achieving FedRAMP, moving into, the Department of War is a natural evolution and a clear relationship between the two. You'll find that most, our our experience has been is that most applications and mission owners for the defense environment are looking to deploy at the IL five level. So achieving that FedRAMP is a key stepping stone for that. Gotcha. Thank you. Alright. So over to you, Ira. As we build on that, how does the scale of AI agents change the identity paradigm? So can agencies just use legacy PAM, privileged access management, as a solution for this? What do you think? Yeah, Sam. So what we're we're hearing and what we're seeing, obviously, is that PAM was really built for humans, and they operate at human speed. Right? So things like you check out a password, you open a session, hopefully, you've got recording going. The thing with an agent is it's it's not your traditional user account. It's not your traditional service account. It really is... It... It's it's own thing, and they act at machine speed. So that could be thousands of threads, thousands of activities happening in seconds. Right? You know, we're already looking at a ratio of, what, 82 machine identities per human, and that was twenty twenty five. That's only going to go higher and higher. Right? So, this is an... Another layer on top of it. And oftentimes, organizations, they start with AI agents. They find they can do all this useful stuff. They find something blocking it from doing it. They give it more permissions. They hard code that in, and it and it just works. And they do the thing until something bad really happens. So, you know, we've even heard, we've seen that, you know, an organization told us in in about nine seconds, an agent found an overprivileged credential. It was trying to do the thing it was setting... You know, told to go do, and, the way it saw to do that was to delete a production database. So, you know, that caused some issues there. Wow. That's unexpected behavior? Yeah. Right. They're sycophantic. They love to please. Right? Yeah. Yeah. That's incredible. So I guess to add to this, you you touched on this a little bit. Maybe you can give us a little bit more of your thoughts on machine identity threats as they accelerate and Yeah. Maybe? Right. So, you know, I think the thing here is, identities and credentials are the attack surface. So you've gotta keep that in mind. Right? Research shows us that when a compromised credential, or a credential compromise occurs, that's when you have a successful breach. And oftentimes, it's a machine related identity that's that's, you know, being compromised more and more than it is a a human falling for this. So, you know, I think the examples here are highlighting that pattern. Right? You can see this trend. It... It's not slowing down. If anything, it's accelerating. And I think the the thing here is a secret should never enter So we're gonna talk a little bit more about that, but, you know, prompt injection has nothing to steal if the agent never holds the credential, and we attribute where that action is coming, from. Great. Thank you. So back to you, Matt. What are we seeing for federal readiness for this new landscape of AI adoption? Fear, uncertainty, and doubt, I I think is the short answer. You know, the the administration, has embraced AI. I'm I'm not gonna say SI, but embraced AI in full throated ways. Right? So we're seeing AI, all the things, push it faster, deploy it in more use cases. You're seeing broad adoption across the department with some sort of narrow, scoped efforts, but just sort of in general, applications that we're hosting across our federal platforms in both FedRAMP and DODFW. It... It's the rare exception at this point that is not AI enabled in some way. And yet you see not a lot of organizations feel comfortable about their security posture with respect to AI security. There's not a lot of direction out there to either federal civilian agencies or defense agencies yet on on secure and strong operating principles. There are developing control sets and limited guidance about this. But, yet that lack of insight and lack of direction and lack of of security framework around AI is not stopping its movement, and that's terrifying. Right? We're seeing that despite the lack of of training for personnel, despite the lack of of clear guardrails, despite the lack of of accurate monitoring and insight, we're still AI ing all the things at the speed of mission. And, you know, mission, there's a lot of demands there in terms of going fast, but it can't be at the cost of security. When we make those trades, it ends badly. Right. Right. Right. Right. Yeah. Well put. So, Ira, could you talk to us a little bit about what organizations need to work towards to start addressing these threats that we were talking about earlier? Yeah. Certainly. So I I think Akeyless is one of the few vendors in the market that's focused on solving for for both paths here. Right? The the platform, that we provide covers, you know, the entire breadth of this. And the reality is that, ultimately, it's it's gonna be a mix of these things. You're gonna have static secrets because you have legacy systems, and it's very hard to adapt those. So, you know you know, we we view this as maybe left being least good and right being most good, but, you know, we understand organizations are are going to have a smattering of these across. Right? And so the thing here to keep in mind is that, as I mentioned with the velocity of what's happening with agents, we wanna offer the same capabilities and secure brokerage, but we can't let those agents have the secrets. That's... You know, just in time is not enough. If thirty seconds is enough to go delete a database in nine, then what else are they gonna do afterwards? Right? So I think you can kind of view this as stages of maturity maybe, but, you know, ultimately, it's a transitional period that, you know, organizations and agencies are gonna have to go through. Thank you. Alright. So let's talk a little bit more on how an agent... So we said even if the agent has legitimate access to the system, what happens if that agent decides to do something it was never supposed to? Yeah. I I think, you know, agents themselves, we can't... We can no longer think of them as just simple chatbots. They are... The models are getting smarter. You have to give them less guidance. They connect to systems. They find ways to connect to systems if they're able to, and, this can mean that they can access things like, you know, sensitive data. They can go make changes autonomously. That's what we want from them. Right? Probabilistic is is the reason they're better than scripts. But they do it at such an immense pace. And we saw, you know, that, what happened at, say, Pocket OS, right, where a cursor AI agent is working on routine tasks, and it's in their staging environment. It says, I've got a credential issue here. Right? I I think I've got an API token over here that I can then use to, fix this problem. Well, the fix was deleting a storage volume that wasn't actually the issue. So the the problem here is that, the agent, you know, it's it's a... Maybe a a toddler with a PhD. And so it sees a a a potential solution. Maybe it thinks that this is the way to go do it, and it finds something with broad permissions, it goes forth, and it does it. And so, you know, this this is basically, something with Pocket OS. We saw that it interrupted a a car rental business. It took down, you know, their their front end. That's money for this organization. And and it wasn't malicious. It was just trying to do the thing that it was tasked with. Right? So that's the challenge here. Autonomous agents are given a task. Maybe you forgot some guardrails around it. Maybe you didn't give it the exact task it should do, and it infers things. And so, you know, what we have to do is, leverage our traditional identity controls that do exist to try to scope this thing down. Of course, we should have Leaf's privilege on these credentials. But once the access is granted, what the agent does with it is, you know, unknowable. Right? So I I think that that's that's the key to kinda hone in on here is we gotta enforce what they're allowed to do, and and ensure that they can't go find other things, to get around the enforcement too. Yeah. I hear you. Recently, I was I was playing in my lab, and I had a secrets manager and had my secrets in there, and the agent wasn't allowed to log in or see any of the secrets. But I had a... An injector inside Kubernetes. So what it did, it found that it couldn't get it from the secret manager. It actually got it from the Kubernetes secret, which was like, woah. How did you get that? Shared cluster, you know, you don't have it locked down. RBAC's not enough. Name pace segregation. I mean, there's any any number of things that we're... We can happen. Yeah. That's a great example. Right? Yeah. Yeah. Cool. So, Ira again, can Akeyless actually control what actions autonomous agents can take in real time? And, of course, I know the answer to that, but I don't like you. Yeah. Spoiler alert. Yes. We can. So I think what what Sam's hinting at there... So, you know, if you think back a couple slides ago, we talked about secretless AI and that the just in time credential is not enough because of the speed at which these things are happening. So, what we do is we also take this a step further. So beyond keeping the credential out of agentic context, we also layer on top of that. So we we kinda sit in the middle. So, you know, we take raw tool calls coming out of a harness... Or raw calls coming out of a harness. We serve as an MCP. And then on the back end, we broker those credentials using the same just in time credentials that we already are storing and managing in Akeyless, and actually make the structured tool calls on behalf of the agent there. Right? And so that's great for the identity aspect, but what we also do is we allow you to call at this time your own agents, whether that be one, two, five, we have quorum, and give them plain text rules on things. So this is, how we can apply intention to the agentic actions. So for example, in some of the previous examples, we've we've talked about with deleting database credentials. We could have a rule that says, hey. No DDL. No destructive activities on a database, And we're going to be able to provide a check on, those agents' actions, as they're doing things and steer their behavior in the correct direction based off of what the rules define. Great. Maybe, Ira, if you can walk us through a practical saying. I I know you... You've touched on that. But Yeah. You have anything else we can talk about here from a rogue AI agent trying to do something that it shouldn't? Right. So, you know, this is this is, maybe the slide I should have mentioned. We should have been on on whenever I was talking about the last thing I was, I was speaking on. But, you know, this is the flow here. Right? So you have your agentic frameworks, whether that's Claude or Cursor or Codex, or something you're hosting in in Bedrock, wherever that may be. And then, you know, we have this, Akeyless gateway, which is a lovely all in one kind of component of our platform that gives you, you know, the ability to go stand these up wherever you can run a container. They egress out to the control plane, and and we can respect your trust boundaries very easily with that. And it also hosts this agentic runtime authority component, that, you know, kinda... As I was mentioning there, it's gonna sit in the middle. So, you know, again, this is this is where we define those... That policy layer layer. We're thinking of, you know, exactly the the types of things. And, I mean, let's be honest, you're probably gonna go ask Claude, what rules should I put on this? And so, you know, you you come up with these rules. You you collaborate. You figure out what those are gonna be, and we're able to do that, in session inspection. One of the things I didn't mention too is, it's not just secrets that we're keeping out of the Gentic context here. We also have output rules. So you can mask sensitive data coming back on those requests as well. And, again, plain text rules say no PII, no credit cards, no socials, things like that, and, it prevents the agent session from ever getting its hands on that data. Which is incredible. So it's not just on the request side. It's also on the response side, which is really, really cool. Yeah. Exactly. Right. Cool. Okay. So, Ira, again, we've talked about securing these AI agents and controlling what they can do, but federal agencies have much broader identity security requirements. So how does Akilas bring all of this together? Thoughts on that? Yeah. Definitely. I I think for federal agencies, the challenge we have here is, they're being faced with incredibly complex environments. They have strict requirements put upon them. And what Akeyless provides is a, you know, unified singular identity security platform. And what I mean by that is Akeyless can really become that front door, whether it's an agent, whether it's a human, whether it's a service account. Right? And so what we can do is we can help you align to the standards, protect and govern those credentials for you, and enforce the zero trust principles that are required, of these environments, through things like least privileged and just in time access. And another thing that I will mention, we didn't totally talk about the platform here, but if you're not familiar, we are a zero knowledge architecture. So we give you all the benefits of being a SaaS while you, as the user and a customer, retain cryptographic control within your own environment through distributed fragments cryptography®, which could be, its own session, in and of itself. So this gives you, you know, centralized visibility auditing, runtime enforcement for AI agents, and it allows you to, you know, exert the control you need over your systems, users, agents, and what they're able to do. So we're all about strengthening the security and accountability and reducing operational complexity for your agency. Excellent. Thanks, Sarah. Matt, back to you. So we've covered what Akeyless brings to federal agencies, but getting commercial technology into government environments comes with its own challenges. So how does Second Front's game warden platform help overcome those barriers and bring these capabilities, to federal customers similar to how Akeyless is doing it? Yeah. Thank you. So Second Front actually, you know, founded with a very similar intention that Ira's just describing, founded originally by marines who realized it was too hard to get the technology they needed where they wanted it, at the speed of of me... Mission and relevance. We do that through a platform as a service offering. We've achieved beachhead authorizations within the Department of of War and within the FedRAMP environment, and we offer a platform as a service with an inherited security model that takes in containerized and virtual, applications, runs them through a security release pipeline that includes hardening functions, generates security evidence, security telemetry, creates an authorization package that can be quickly adjudicated by a government authorizing official either through the FedRAMP PMO or through an agency sponsor or on the Department of War side, through an authorizing official. By by by offering that, what we do is we tend to to obviate the undifferentiated heavy lifting. A lot of the compliance requirements that take engineering focus away from making an exquisite application or exquisite software experience, we do that by providing our platform. So we're handling things like, ingress and egress, identity management, security security operations center, continuous monitoring, those kinds of things. And with that, we're able to reduce the control burden on individual applications, you know, 70, 80%, of the way. So when you think back to those early slides, we were talking about the controls read requirements that apply to the high walls of of government software authorization. We obviate the vast majority of that, take that, non value driving activity away from folks so that they can focus on on delivering a quality experience. Yeah. Thanks for that. Okay. Well, thank you both. And, as we're we're speaking here, it's clear that AI adoption is just gonna continue to accelerate. But giving credentials in the hands of these agents and protecting that is not enough. I think enforcing this runtime authority that we've been talking about is is paramount as well. So we have a few minutes left here, folks, on the call. So if you want to post your questions, please go ahead and do so in the q and a. I do have a couple here. First one, Ivan's asking about, the recording. Yes. We will have recording. I have another question here for Ira, and this is about audit purposes. So can you trace an agent's activity back to the original user request and policy decision? Yeah. That's end to end within, you know, Akeyless. So the audit trail comes from that remote MCP that we were talking about. Right? So you get user prompt, you get the classified intent, a policy verdict, and, you know, the ephemeral sessions activities on that target. So... You know? And users sign in to our platform and provide that access via, you know, your standard SSO or SAML or OIDC flow. So every agent action carries that user attribution, and it's not a, you know, shared service account here. So, and, of course, all of that flows up to your SIEM and system of record tooling. Excellent. Another question. And I think you touched on this. Maybe if you can elaborate a little bit more, Ira. So someone's saying we're looking at AI agents that may need access to several systems to complete a task. How do we give them enough access to work without creating overly broad permissions? Yeah. So Akeyless as a platform lends itself to this, whether it's an agent or a machine identity. We're able to, you know, generate just in time ephemeral credentials on several targets. Where it's not supported, we can make rotated credentials. Or even if it's a static secret that we need to lend out, we can lock that from changes and check it out to different targets. So, you know, you're just kind of reasonable concept of a target within, Akeyless, and then the item that defines, the task or the secret type that's gonna be used on that target, you can gate and control least privilege there by scoping down the creation statements and say databases, for example, or the permissions that are allowed out in your cloud service provider. And so, you know, each target gets its own scoped role, and access to five systems means that we can have, very granular, different access to each system there. So, on top of that, with that defining maybe the ceiling of the activity, runtime authority then narrows that down to the subset of allowed activities, that that permission grants. Yeah. Thanks for that. I think we got a couple more. If an agent has legitimate access to a system but tries to take an action outside its assigned task, how is that detected and blocked? So I guess it's tying into runtime authority once again, Ira, if you if you wanna touch on that. Yeah. Certainly. So I think that's kind of the gap maybe I was I was hinting at there that, even though we've got granular RBAC and access and least privilege granted, we're gonna check every action that an agent requests against policy, and it's declared intent before it actually reaches that target. And so, you know, in... For example, my own, demo environment, I have an agent that can read a database master password stored in AWS secrets manager. And so IAM by itself in that policy would have allowed it. However, runtime authorities policy says, you can't assume that role. You can't go try to find credentials, do exploration. And then, you know, as we mentioned earlier, things like output can then be masked and blocked back from entering agentic context, if it is able to make certain queries to databases and things like that. Great. Last one I see here, unless anybody wants to add more, and this is for Ira as well. If we wanna move AI agents to production, what identity and access controls should already be in place before we do that? Yeah. So I think, you know, one of the things that you're going to want to have here is, as you mentioned in your your environment, you don't want credentials, in agent code, in config, an MCP file, places where agents can reach it. Right? So, you know, that's really, I guess, the floor, the the barrier to entry here. And then, you know, tying agentic action back to the user itself. So you don't wanna you don't wanna be leveraging shared service accounts to kick these things off. And then, of course, short lived revocable credentials, on any of these targets where, you know, where it lends itself to that. And then, you know, obviously, we've gotta have great audit controls. We've gotta be able to revoke those identities very quickly. And then we need, you know, policy in force, for anything that's gonna touch production or sensitive data. So, you know, you've gotta basically get ahead of this, and you've gotta make sure you've got a really solid foundation because retrofitting all of these concepts on the back end is going to be like, you know, building the car while it's in motion. Right? Very difficult. Well said. Well said. Alright. Well, I don't see any more questions. So once again, thank you, Ira. Thank you, Matt. Really appreciate your time, and thank you everybody for being on the call. Once again, this is gonna be... Or is being recorded, you'll get the recording after. So appreciate you, and take care, everybody. Thanks so much. Thanks a lot.