Name: Automated Red Teaming: Attack Path Chaining in the AI Era
Uploaded: 2026-09-29T10:01:23.475Z
Duration: 33 min 8 s
Description: Automated Red Teaming: Attack Path Chaining in the AI Era

Transcript for "Automated Red Teaming: Attack Path Chaining in the AI Era": Hello. Hello, everyone, and thank you so much for joining. Maybe let's give it another few seconds. But otherwise, yeah, you know what? Let's get started. So thank you again for joining us. You know, whether you're on a red team or if you're running a security operation center, you're managing exposures, or maybe you're just a cyber security enthusiast, maybe you're a filigree enthusiast, I don't know. Today will be useful for you, I hope. My name is Pierre Loic. I'm a senior product marketing manager at Filigree. I have about eight years of experience in the tech sector in IT and in security, and, I've been at Filigree for about a year now. And today, I'm joined by Sebastien who leads, our product management team. I don't know, Sab, if you want to introduce yourself very quickly. Yeah. Sure. Hi, everyone. So, yeah, I'm Sebastian, the the product manager for Pen Ivy. Yeah, a few words about the background. So before John Filigran, almost two years ago, I worked in the cybersecurity, in both the blue side in the SOC in build and the run, and switching the red one, as a pen tester and close the loop, when, yeah, I led the purple team and the CSIRT, for a French MSP. So yeah. Hi, everyone. Amazing. Thank you so much, Seth, for that. And, yeah, it's gonna be a very good interesting session. On the agenda today, we are going to really go through a few different things. So we only have thirty minutes in front of us, so it's gonna be a short and sweet webinar, but we do wanna make it as practical as possible, to give you context, to give you a snapshot of what we've discovered and, how we're tackling our problems at Filigree here. And at the end, we'll, of course, be going through some demos, which is gonna be the really exciting parts of this, of this webinar. So a few housekeeping before we start. This will be recorded, so you can watch the replay or send it to anybody you think would be interested. You can also put your questions in the q and a box that's dedicated for that. If you have questions that you wanna ask live, and we can answer them live. And also you will find a few documents, I believe, in the interface, if you're interested in reading a bit more about what we're seeing here today. All right, so let's jump straight into it. And to start this, I really like to help frame, with the cybersecurity sec, what the cybersecurity sector is facing today with something called the cybersecurity paradox. I'm sure some of you are already a bit familiar with it And really, it's about how we have, you know, the general investment in cybersecurity is growing exponentially from tools, solutions, team processes, and these really have been increasing fast and fast and fast. And by all accounts, we should be more protected than ever, and yet we have a huge number of re of breaches still happening. So spending is up, and yet breaches are also, you know, up. So we're reaching a point where we're starting to ask questions. We're spending all this money, but we're not any safer. So what's what's happening here? And so, of course, there are many reasons about this, and so we just have to focus on a few. And of course, it's related to how, you know, the modern threat landscape is changing extremely fast. You know, the attackers are attacking faster than ever, and in much more sophisticated manner. The stacks that we're using to protect ourselves are getting bigger and bigger and more and more expensive and harder and harder to manage. And also often those tools that we use, those controls are assumed to work but not always proven. So you might have your ADRs, your firewall, your interface, but can you actually prove that they are preventing attacks? And all that to say that these attackers are working continuously, but the way we defend ourselves, is not matching that level of attack. So, what you can also see is that AI is making everything even worse. Right? So, AI grew the back dog instead of shrinking it. And you can see how this data from from volcic kind of really helps, put it into perspective. You know, you can see how, for example, Chrome CVEs by May of this year, have already surpassed everything we've seen just in 2023. And this pattern, we see it everywhere. And so discovery is actually outpacing our capability to remediate these issues. So it's not because we can see more CVEs that we can fix them all. And actually, that's the problem here is that the gap between what we know and what we're able to fix keeps on widening. And we've done our own research here at Finnegan. You know, we are this open source company and we pride ourselves in transparency. And so we wanted to conduct our own third party research. And so we have this state of threat management report we released earlier this year. And of course, we've, you know, uncovered some really great insights and two of them really stand out. You can think of the exposure gaps, right? So how, even though we have all these tools and we have all this visibility from threat intelligence and from the different feeds that we use, people are still not very confident about the way that they are looking at the attack surface. In fact, 93% don't even have or feel like they don't have an accurate view of the attack surface. And, you could also talk about the prioritization gap. Right? So there's a real problem here and the fact that it's it's not just about having the visibility into your threat landscape. It's how you prioritize the overwhelming amount of vulnerabilities and signals that you are faced with every day. And so this fits really with the, with the cybersecurity paradox. Right? You might have the tools. You might have the visibility, but yet you're unable to manage the amount of threats that are getting through. And so, if we have to summarize this, right, in a very simple way, we can talk about the mindset shift of the the the SOC leadership or what they're facing today. And before they used to ask this question of, okay, do we have the tools to be secure? You know? Do you have the products for every category of your security landscape? Do you do you need to check all the boxes, etcetera, etcetera? But really, the question now is about proof. And the question you need to ask yourself is can we prove that we're secure? Right? Whether you're a CSO board, an auditor, an insurer, you need evidence that your controls are actually working, not a list of what you've deployed and things that you have right now running your organization. So let's jump into the second part now of of this presentation really, which is about what it means for the offensive security teams, right, because this impacts everyone. But specifically we need to talk about those whose job it is to actually test and, you know, make sure you get this proof and get that evidence. Of course, there are many problems that we can discuss, but let's jump into a few of them. The first one is a very common one that that that most, people watching today will probably be aware of, and that's that attackers don't really use one weakness. Right? They will chain them. Right? They will chain several weaknesses. They don't, you know, plan on betting on this one big vulnerability. Rather, they'll combine several small ones, which each opening the door to the next until they get to the objective. And so if you look at these findings one by one, for example, on the one that you see, none of them are maybe rated critical. You know, phishing or password or an accounts, etcetera, they are all medium or maybe low findings and they also in the backlog, but together they will create that straight line to your main objective. And so this is the problem because testing today is starting to miss this more and more. You know, testing isn't built to see this right now. The scripted test, if you use scripted test, they don't run the same steps every time and should they'll find those more weaknesses, but not the entire path. And on the other hand, yes, you might have manual red teams who are conducting kind of deep pen testing scenarios, etcetera, but they will follow a path, but only a handful time a year, maybe once or twice, depending on budget and time that you have. Right? And so we have to ask ourselves why this matters now, because the question is that the test from last year will say very little about today or test even from last month might say very little about today. So the point here is that then the point in time testing really can't keep up anymore because your environment is changing, whether you're installing new apps, you are having new rules, or, you know, you're giving new access, etcetera, etcetera. The pen tests that you're running are capturing one moment, and that moment is gone practically immediately in today's landscape. And in part, it's because AI is compressing the attackers timeline. So attackers already use AI to do reconnaissance faster. They write phishing at scale. They adapt as they go. And so the window that you have to detect and respond actually keeps on shrinking. And the last part of auditors, they expect recent evidence to date. They are gonna stop asking, okay, when was your last pen test? And they are starting to ask, when did you last do your pen test? And and what did you find? Can you prove that you're safe against these specific explanations? So it really is about, being able to see how you can test against the the current periodic system. And so this is also highlighted in our, reports that we've, that we release the state of fit management. So this is proven with 88% saying that periodic assessments alone, they can't keep up with this changing environment, but also 82% saying that they are slowed down by manual work when they are spotting urgent risks. All that to say that, you know, if we look at the red and purple team challenges that they're facing today, there's this uncomfortable truth that they're faced with that most organizations are testing much less of their real attack surface than they actually believe. Firstly, the first challenge that simulations don't actually behave like attackers, so most simulations run the same way, whatever they find. So if a step discovers a credential or, the scenario doesn't necessarily use that, or if a real attacker would would use that scenario, they would pivot immediately. But the the scripted test cannot provide that. So teams end up validating against scenarios that don't really look like real attacks. Right? The second challenge is that we are still facing kind of pass or fail system with very little explanation. So how often can you actually run a scenario that tells you exactly what happened and why a path worked or which step actually made the difference? Today, we still get the green or red pass or fail system. And so you end up triaging every single weakness instead of actually fixing the one that really matters. And the last point, of course, is is one that we mentioned is the the red team's capacity, doesn't always scale. Building a multi step attack, you know, by hand takes a lot of skilled people and time, and budgets. And so your coverage ends up following, well, that limitation instead of your actual risks. So let's talk about then quickly the shift that we required into this landscape in order to make this work. So the the current approach is quite simple the way to think about this. The we need to test the ways that attackers work, but not just that, it needs to be continuous and it needs to be test step by step. So first, steth testing can't be a snapshot anymore. It has to run continuously, so it picks up changes as they happen. Second, scenarios have to react the way that they discover. So if a step finds a credential, the next step should use that credential just like an attacker would, and they would pivot in the same realistic way. And the third, automation should do the repetitive execution. Sure. But then people should still set up the strategy and decide what matters instead of running the the same steps by hand. And then finally, coverage should stop depending on team sizes. Again, as we mentioned, we need to you know, you shouldn't need twice the people to test twice the path. That's just not sustainable anymore. So one thing that is very important to to highlight here is that manual testing is still very, very relevant. This isn't about replacing red teams. I want to be very clear because this is a question that we get lost. We're not saying that red teams are obsolete or anything like that. To the contrary, it's just the scale of it that is the problem. Manual engagement is still super important for human creativity and physical testing that it can only bring, that no digital simulation will able be able to replicate. But they are limited by expertise, by budget, and by scheduling. Right? And this is where we need to to to find the the fix. And so the takeaway here is that automation should handle the repeatable high frequency testing, and then your expert should then focus on the novel techniques and the creative attacks and the judgment calls that they need to make, which is what they're hired for. And then that feeds back into what should be tested continuously. So, yes, again, if you have your periodic testing, they need to be complemented by some way to automate a continuous way to test, those, what you've already found. So now let's jump into the last part, and I promise we'll jump into the demo, which is about what we do at OpenAV, with attack chain. And but before we even talk about OpenAV, I imagine most of you are familiar with Filigram or maybe Open CTI, but OpenAV isn't a standalone product. It doesn't work in isolation. It's part of the wider XCN platform that we have. Right? It's is goes alongside OpenCI and alongside also OpenCRQ for cyber risk quantification, where all these can work together. OpenCLI feeds OpenAD the threat intel, so you're testing in techniques that matter and, the result in flow back at security posture, etcetera, to OpenCLI or even to OpenCLI that is released very soon. So you can link the results to actual business risk. And then all of this is supported by the agentic layer, which we call XTM one, which we'll get to see, in a second during the demo. And so OpenAV, really, then if we zoom in specifically OpenAV, it's based on the four core pillars that are required of offensive security testing, which is a breach in attack simulation, which allows you to test your technical landscape, but also tabletop and crisis. It's like you test human readiness. And then we can automate this to ensure that you can then, you know, build proper automated fantastic scenarios, and then, of course, allow you to do continuous exposure validation over time, autonomously. And all these work together in a single platform, which is transparent. So there's no black box, you you know, which is very important when you're running a text in production. And it's built for your environment, so you can work with your current stack and you can bring your own DER. It works with the tool that you have. And, of course, it's prioritized by threat intelligence. So, again, it's it's it's about prioritizing what what matters most, and you can use your CTI in order to make sure that that happens. So, attack chaining, and I will just introduce it before before, Seth actually shows you how it looks like, but really attack chaining is the feature of OpenAV that provides a live automated fully adaptive attack path engine designed for continuous red team. And it orchestrates the entire chaining scenarios, manually or autonomously using conditional logic. And so for example, as an example, right, you you would run an action, and then the action finds an output that becomes a finding, then finding then automatically picks the next step, and and creates a next action based on what it discovers. And then that loops continuously until it gets to its objective, and all of this then is visible on a live map. And so what you get really in practice is that you can then build your own conditional logic. You can see the attack unfold live in front of your very eyes, and then you can spot the choke points that you need to remediate where one fix will break the entire path. And, And, of course, like any one of these tools, you have the guardrails available to make sure that, you know, you can fix as it goes. And so this is two there's two ways to run this. Right? There's the operator led and then these autonomous led. So operator is you build a logic, you have full control. Just make sure that you repeat this that you want at scale, and you can put as much configuration as you want. Or you can go fully autonomous. And by fully, I mean, fully. Well, the agents, they do the entire thing. Right? They are exploring the entire path that nobody has scripted, and they will take care of the entire, scenario for you. And, of course, there's always it's kind of a spectrum. Right? In between, you can decide how much you mean the group you want to incorporate, how much you want to control your agents, and how much, you want to manually configure yourself. And listen, this is pretty much it. I will not pass it on to the demos for for for Seb so we can actually show you how this is working. But, you know, just to finalize, you know, more path tested, more gaps found sooner. You can fix matters and you have the evidence in your hands, with this platform. But again, I just said a lot of sub words. You just saw lots of slides. This is all conceptual, and so you actually see it in practice. So I'm gonna pass it to Seb who is actually gonna show you what we just discussed, in a real demo scenario. So, Seb, over to you. Thanks, I will share my screen. If you can, stop to share the the slides. K. That should be good. Yep. Perfect. Okay. So, yeah, let's start by a quick introduction, within OpenAV, from the dashboards, and then we will switch, in two part of the demo. So the first one will be more about internal assessment, like activity activity assessment with command vulnerability, like, bursting, as a posting, and, yeah, any kind of those elements that can be also really relevant for the company about the share permissions, assessment, which is a a huge works for a lot of companies. And, this kind of scenario, which will be shared by, the hub to the community can be a good entry point for the those kind of people. So if you're part of them, yeah, you will have access to that sort of extent hub. So, yeah, quick introduction on the dashboards. So as you can see, we have a lot of, a lot of elements. We have a, dot here widget where we have a critical exposure score, which is based on the wall, elements you have in the platform based on the time period you're working on. So as you can see there, it's three months, but you can update it, of course. And we have the score, which is based on the different expectation results you have. So that's based on the vulnerability, which mainly is based on the CVE, the prevention. So is your idea prevents the payload to be run, the detection. So is your EDR or SIM let the payload be run, but you catch it. An alert, you know, it creates an alert and you catch it. And manual. So, yeah, if you have any human response, on the. So, yeah, as you can see, that's not really, really good for us today, but it's fine. That's a demo environment, so don't queries. So, yeah, free words. We have a lot of widgets. Everything there is customizable at 100%. It's really part of the data, and I I think it's worth to mention it so you can do your own. And those with the dashboard can be con contextualized, whether it's in the scenario, the simulation, or any part of the elements you're working in OpenAV. Of course, keeps shutting evolves, so we have different key keychain, attacks reported, like Atlas for the AI assessment, attack, which is the basic one, the enterprise frameworks mainly, and we have others like OTICS, cloud, and so on. So, yeah, there is a lot of thing. I will not take too much time about that just to highlight, yeah, those elements. And then, yeah, I can switch on, the second part, which is an internal assessment. So let me share this one. Up. Should be good. Yep. So this is what we call an attack pass today in OpenAV. That's really, a huge feature, a huge product, a huge version we released few times ago. So it's part of the history, basically. And the attack path, the attack training is really the core of that history. So how it works, it's, basically, today, people like testers, purple teamers, and bad guys as well think about, events. So that's why we switched from the time based approach we had before to the event based now. So, yeah, how it works is quite simple. Each time, you have a condition which is fulfilled, you have an action. So you can say, okay. Each time you find the port four four five, then do a SMB scan. Each time you have the port three eight nine or six three six, then make any assessment on the LDAP side. That also can be applied on the world scope we have, meaning that can be on the share permissions. So meaning if you have any shares with the right end of the the hidden device permissions, you can leverage them through an event that can be on part of the user, the username. So it's if a user a username start by admin or something like that, you can do specific stuff. Or each time you find a credential, then try to use the credential you found to authenticate on my IDP, like Azure, my active directory, but also my SSR. Why not? So, yeah, that's really, a new approach we have, and the attack pass is there to show what really happened. So that means we there have an attack pass which show us we have, a network scan done by the the capability we have. So we have a lot of concept that will not take too much time because, we don't have too much time today. So we can leverage, network mapping like end map injectors to check any postcard. Then we have some open ports as you can see there on the Winifel machine. So we have the open port, the four four five. If I click on see on it, I can see this one trigger the SMB share listing. So I I reckon use the attack pass as a denture point to browse what really happened, you know, on my simulation. And I can see also in the overview what will happen. I can see the number of share I had, with the permissions. So I can see, on that IP, the old share is accessible with the real and wide access. I had some, files captured inside them, so I can see, yeah, there is a secret.ps1 share, which is, p s1five, sorry, which is inside the share. So I can really see what happened. And plus of that, which is something really worth to mention, it is we have, the agenda to explain it. The idea is not only to say there is misconfiguration, there is vulnerabilities. It's also to take care about the defensive dimension on it. So what does that mean? It's quite simple. If you have, a SMB share listing done on your wall networks and you don't have any alerts, and the guy, leveraged the files to, you know, get the files to check if there is any secrets inside the the the the share you had in the files and the shares, and you don't have any alerts. That means yeah. Today, everyone can do it without any detection in your side, and it may be worse to take some time to look on that. That's what we call the choke points. So today, choke points are there to help you focus on what really matters based on the asset, which is set on the asset side directly and can be, of course, be updated as you want, but also based on number of the findings. Meaning, if your critical asset is with the number of finding, that will ponder the criticism you have. Okay? So choke points are there to say, okay. In my case, the winter cell machine seems to the top asset I have to focus on. Okay? So yeah. And, basically, you can use the search bar. You have everything you want. You can enter the the value you want. You can switch the table view. You can export it. You can leverage it to switch in the attack pass view. So, yeah, this is the attack pass, and I think, it can really, yeah, take a dedicated decision just to go in the wall attack pass capabilities. We each time of the prevention detection results. So meaning, is my cross track prevent or detect that, net exec assembly sharing information? Seems he didn't. So, yeah, I can see the terminal view. So there I can see there is, yeah, some shares as if I was in a, you know, in the shell, basically. The findings, it creates in a denomination. I can do to improve that. Okay. So yeah. Also part which is really, really a huge one, it's about AI because, yeah, today, everything is AI as you noticed. So let's switch in how can you do, a chance scenario maybe with without any knowledge in OpenAV because, yeah, if I quickly show you how it look likes, that's a pretty simple scenario. You know, the one I just showed before. It's basically, you check if you have any port like SMB, then you you check, the SMB. You have the shells. You check if there is any read permission. And if yes, then you check the files within the chairs, and then you you you check the files. You know? Same for the LDAP. If you have any LDAP open, then you check it. Cabo hosting and as a posting. So that's a quick simple one, but it can be complex at some time because you can really customize as you want, which is really a strength we have. You can do really what you want. But, yeah, if you don't want to be, to take too much time, you know, to understand how create a scenario you want, There is also AI capabilities to help on that, through extend one. So that's, that's a simple, example I did. So I just sent a prompt. You know? So you have you you you can just say, okay. I don't know how to build a scenario in OpenAV. I'm not an OpenAV yet, but I really would like to create a scenario around exfiltration, which concern, which could concern public sector to be sure I can detect it. So, yeah, let's do that, please. And then yeah. I I will do it for you. We'll ask some question. You can see the decision timeline there so you can see the different step on it. So he will ask some question at some time, like, okay. On which asset do you wanna have a need? Okay. And, and you will have, on the right side, as you can see there, the iteration within it. So you already created the logic page with the wall, the wall configuration of this, scenario. And then as it's in a autonomous mode, it will automatically run it. You can just build it through AI. You can also build and run it through AI, and AI will learn based on the statue you have. So for example, AI have an error, but AI will learn from that error to create a new payload, and, yeah, improve that part. So we have the two options today. And, yeah, I think that's, that good to to mention that. And, okay, I think it could be worth to also explain a bit about the phishing capability we have. There is so we can leverage the phishing within the scenario we can, do with some. So that means if you wanna create a phishing assessment and then check any credential, shared by the people in the company, leverage those credentials, checking the credentials are valid. You know? You could totally create efficient campaign. So that's basically, common today. You know? But you could say, okay. Let's say I wanna do a Office 365, postal. Okay. Please generate it with AI. It can be, a VPN, a corporate VPN portal. It can be a GCP one, a specific portal you want to know, and then AI will automatically, create it for you, as you can see. And then you will have the preview, and you can set, okay. Do you wanna redirect some, to a specific URL? Okay. Do do I accept that? Yeah. Let's say I do. So yeah. Which product URL after the submit? Do I want to capture the submitted data, the password? You can do or don't based on what you want, and you should do capture the password. You can leverage them through an event to say, each time I have a credential, then I use it to my IDP, my Azure, or my o two sixty five portal to check if that's relevant to credential I had. Okay? And, yeah, maybe last thing, because I think that that could be worth to mention it as well. Oh, if I go back there, once you have an attack pass like that, you know, and you want iterate on it or you just want to take and pick up some action on that, you can generate the reports. The report can be previewed in, the OpenAV platform as an HTML one can be exported with a HTML or PDF up to you. And you can directly see, yeah, the KPI you have based on that scenario, the different kitchen coverage you add. So you can see green is good, but, redispad as always. The breakdown you have, which security scope is covered. You know? And you have some datas. And once again, as always in the products and open a v, of course, everything is customizable, so you can make your own. So yeah. Yep. Yogurt Peel. There we go. So thank you so much, Ted. It's so much better to see it actually, in action, see the real practical thing. And so we are out of time, but I think if you wanna stick around, people, we can answer just a few questions. We have three questions that actually came in. The first one is what's the difference between weakness and an attack path? So I'm happy. to take a look. I don't know if you wanna add colors. Actually, not. it said from from your perspective. That's. So, basically, weakness is part of an attack pass in my opinion, because we can if we take weakness, like, I don't know, a web, a web liabilities, you know, a web weaknesses, like, I don't know, LFI or what you want. It can be leverage to an event to say, okay. Each time you found LFI, then you can assess the LFI endpoint you found to check, yeah, the different payloads and how it's detected or not detected by your wife or your sim or boss, and then how to improve that, how to mitigate that, how to detect it. You know? So for me, the weakness is part of the attack pass that can be leveraged to the events to build the attack pass. Okay. Perfect. There's another one that came in. I'll let you answer it, Seb, because it's I think it's more directed towards you. And let me know if it makes sense. So what are the restrictions in terms of scope or domain? For example, can I only run internal AD assessments with my chains? Yeah. That's a really good question because today, we basically focus only on one aspect because we didn't add too much time, but it was to mention the the training, the attack training, and the call we have, it's not relevant to only one scope like internal assessment, but it can be really modular. And that means we can mix the different secure domain we can cover. So today in OpenIV, we can have some AI security domain, cloud, email, networks, endpoint, and and others like OTSS and others, web, of course. So that means you can leverage some web vulnerability, switch to the emails, to the networks, and others to to combine through a realistic and huge scenario, which could make sense for some company to be realist as some attackers could do in the wild. So yeah. Perfect. Thank you. And there's the last one, which is just, is it free open source? And, I'll interpret that question. Is it part of a community edition? And the answer is no, this is an enterprise edition feature, the attack chaining, which is specifically what, said demoed. You know, we still have the open the free open source community edition available for everyone that you can download now and use. And the feature you saw is what we consider an advanced mature feature for companies who want to be a bit more more advanced. And so that's one of those features that are included in the enterprise edition. But we do have free trials, and there's a full HTM platform free trial that you can that you can run. And I believe it's even in the docs section in the internal documentation, where you can kick off your free trial and try it out for yourself if you want. And I think we are out of time. We are even a bit over time. So thank you everyone for for sticking around. Thank you, Seth, for for the great demo. And, I yeah. Thanks a lot, everybody, and have a great rest of the of your day and week. Thanks, Pierre. Thanks, everyone. Cheers all. Cheers.