Video: Threat Briefing | AI Threats: What’s Hype, What’s Real, and How to Respond | Duration: 3024s | Summary: Threat Briefing | AI Threats: What’s Hype, What’s Real, and How to Respond | Chapters: Welcome and Introduction (16.495s), Unpacking AI Hype (96.505s), Cybercriminal Hype Tactics (215.145s), AI and Misinformation (924.16s), AI Malware Maturity Model (1008.575s), AI Malware Evolution (1390.73s), Defending Against AI Malware (1641.02s), AI Impact Summary (1845.935s), AI Malware Defenses (1929.145s), Malware Model Tracking (2103.105s), AI in Attacks (2225.65s), AI Security Challenges (2446.565s), Concluding Q&A Session (2976.325s)
Transcript for "Threat Briefing | AI Threats: What’s Hype, What’s Real, and How to Respond":
Hey, everyone. Thank you for joining. We'll give everyone just a couple more minutes to make sure everyone is let in, and then we will kick off. Hey. Good morning again. Thank you for joining us this morning. We're here to talk about AI hype, separating fact from fiction. So really excited to be able to present for you all today. Just a bit of housekeeping before we jump into the webinar content. This session is being recorded, and we will send out that recording hopefully within the next day so that you all will be able to have that. If you have any q and a, please enter that in the box to the right of your screen, and we will address that at the end of the presentation. So now I'll turn it over to TJ to do a quick introduction, and then we'll get started. Hello. My name is TJ Nelson. I am the senior director for technical analysis for the incident group here at Recorded Future, and I'm excited to talk to you today. Megan? Great. Thanks. And I am a senior strategic threat intelligence analyst with INSICT Group as well. So if we could get the slides up on the screen. Great. So AI, as I'm sure we're all pretty aware, is a new technology with the potential to change our lives in possibly science fiction worthy ways. And I think we've all seen ways that AI, in the last couple of years, has maybe surpassed our expectations and lived up to that hype. And we've also seen maybe headlines that were way overblown and, of course, a lot of doomsday predictions around this technology. And as defenders, this makes it really difficult to separate out what threats are real and what threats we should be preparing for and what threats are maybe more in the science fiction category. So today, TJ and I are going to tell you give you some strategies for parsing these headlines and figuring out what's hype and what's real. So we're going to start by unpacking hype itself. So these narratives don't just appear in a vacuum. They're often intentionally cultivated and spread in the media ecosystem. And we're going to focus a lot in particular about how criminals generate hype in order to advance their objectives. But we'll also talk about the broader broader ecosystem as well, how journalists, cyber threat researchers, and security vendors also contribute to this environment of hype. Next, I'll turn it over to TJ who will go through the artificial intelligence malware maturity model, which is a framework to help evaluate malware or claims of malware to assess what's real and what's fiction. So it gives you a sort of benchmark against which to measure malware maturity. And then we'll look at how recorded feature can help you analyze malware and other evidence to really gain a sense of that ground truth. So we're going to talk about hype and how criminals are very aware of their public perception. And this is not something that's new to the cyber age. This has been happening for hundreds of years. So I've got two criminals here. You may recognize some of them. The first is Edward Teach, also known as his alias Blackbeard. He was a pirate in the eighteenth century, and he used the alias Blackbeard very intentionally. He cultivated that in the media of his day in order to make sure that his reputation preceded him when he went out raiding. He wanted people to know that Blackbeard was coming and to surrender immediately and give up their treasure if he attacked their ship. You can imagine how that might benefit a pirate. The man on the other side is John Dillinger, famous bank robber of the twentieth century. He cultivated a different persona, instead cultivating media attention in order to promote this Robin Hood esque figure. And he did this in order to attract public sympathy, which if your face is on a lot of wanted posters around the country, you wanna make sure that people treat seeing you as a bit more of a celebrity sighting than a reason to call the FBI. And cybercriminals cultivate hype in very much the same way, and we're gonna talk about direct and indirect ways cybercriminals do this. The first way is directly contacting researchers and journalists, so acting as their own PR firm. Cyber threat actors related to or associated with the group DragonForce, for instance, directly contacted a BBC journalist this summer providing them proof of their exploits against British retailers M and S and Harrods. And they did this not only to claim credit for the attack, but also to help burnish their brand. They wanted the re the journalists to publish stories about them so that the public associated them with these really high profile attacks. We've also seen criminals maintain FAQ pages and set up dedicated telegram channels that are directly targeting journalists. So, essentially, making sure that journalists have a place to go in order to get the threat actor's version of the story. Finally, we've also seen them release public statements on these FAQ pages as well as indirect outreach that are intended to set the record straight on how the criminals see themselves. And in a notable example, we saw the group Alafvi reach out to journalists at Bleeping Computer and other cyber outlets for what they saw as inaccurate reporting on their attacks on MGM Studios. They thought that what was being reported in the mainstream media was not promoting them in the best possible light, and they had an operational reason for wanting to correct this. It was because they were worried that if the public believed this, they wouldn't take them as seriously when they tried to carry out their extortion attacks. And they were afraid that this would lead to fewer fewer payouts in the event of a ransomware attack. So criminals also influence media coverage indirectly as well, and this is largely through some of their activity on criminal forums and social media channels where criminals operate. They know that cybersecurity researchers, they know that journalists as well as law enforcement are all hanging out on these locations, essentially observing what the criminals are doing. And so sometimes their activity will be rather performative to try to attract attention. In some of these ways, they may be claiming high profile victims or headline grabbing targets in order to prompt journalists and researchers to cover this and to write about it. We saw this with Lockbit claiming, for example, to high to have hacked the Federal Reserve. Obviously, this would be a very big deal if this had happened to be true, so this gained a lot of headlines for the claim. We've also most recently seen a group going by zero APT who claimed over 200 high profile victims all at once, generating a ton of headlines about their activity. Now a lot of these claims actually proved to be false by the time researchers were actually able to dig into some of the claims and parse through the evidence for what they were saying. But researchers found that the while the claims are false, the ransomware associated with this group is in fact effective. And so the theory is that they were using these claims in order to have us talk about them, in order to get these headlines so that when they do carry out attacks, their victims will know who they are. Finally, we've also seen them use branding logos and personas. So in addition to providing evidence of MNS and Harrods, for example, Dragonforce also sent a logo of a really fierce looking dragon and providing their name and some of their branding to the journalist to kind of help amplify their persona. We've also seen this with the group claiming to be going by shiny lapses hunters, which brings together three high profile criminal names and insinuates that they are some sort of amalgamation of scattered spider, lapses, and shiny hunters. Despite researchers, particularly some of recorded futures research looking at this, indicating that this is largely an effort by shiny hunters and not an amalgamation of multiple groups. So we've talked a bit about how criminals influence, but they're also targeting a couple of different audiences as well. And the first one is an ego boost. Criminals have big egos. They want to feel like they are the best hacker in the room. They also want to try to amplify their own clout among other cybercriminals. So that's an important motivation for some of this activity. But there are more operational reasons as well. The first is targeting the public. So you and me, potential victims of ransomware extortion attacks. And the goal for this is to spread fear and intimidation. So very much like Blackbeard was doing in the eighteenth century, they want their reputations to precede them when they conduct extortion claims. And, operationally, spreading fear and doubt is pretty valuable to make things more difficult for cyber defenders to figure out what's actually happening. But at the same time, it also contributes to helping ransomware operators establish a sort of trust with their victims. And so this is a paradox that researchers have termed the ransomware trust paradox, which essentially states that if criminals want their victims to send them millions of dollars in order to release their data, ransomware victims need to believe that that payment will be effective. So criminals do a lot of put in a lot of effort in order to ironically establish trust to seem as credible actors that if you send them the payment, they will, in fact, release your data. And the final set of audience is actually other criminals, and the goal here is to help advertise their criminal malware and ransomware as a service products. In the last couple of years, the cybercriminal ecosystem has gotten much more competitive, driving criminals to attempt more innovative as a service products. So we've seen this not just with adding in features to ransomware as a service and other platforms, but also in monetizing different aspects of the cyber kill chain as well. So criminals are getting increasingly specialized and increasingly innovative in how they're offering these as a service products. And this is really where we start to see criminals talk about AI. The targets for some of the hype that they are promoting about AI is not necessarily the public, at least at this point, but we're seeing them advertise this as adding in new and innovative features, actually kinda much like a lot of legitimate products are doing so now. And so I wanna dig into some of the different AI products that criminals are promoting. So we saw a huge spike in primarily AI how to guides kind of shortly around the or after ChatGPT really started to take off. We saw this spike in how to guides where criminals were just sharing knowledge, often selling how to books to other criminals for how to use these tools in malicious ways. But we also saw criminals put out uncensored large language models that were specifically designed for criminal use. Some of these earlier tools were more fraudulent than actually legitimate. Criminals would put out a product that was basically chat GPT or some other commercial large language model with a criminal branding on it, claiming that it is a specially designed criminal LLM. Obviously, it didn't work as intended. But these offerings have gotten more sophisticated, and in 02/2025, we've seen at least three new criminally focused LLMs that can operate phishing emails, help you make malware, and essentially do other things that most commercial and open source tools won't do because of their safety guardrails. Of course, the creators of these products claim that they are much more sophisticated than your average commercial tool, but an in-depth analysis shows that they actually operate more as on the same level. So certainly still something to take seriously, but not necessarily a more advanced tool than what's commonly available. We've also seen them bake into bake LLMs into different malware and as a service offerings. And especially phishing as a service has been has found really effective ways to incorporate LLMs. We have SpamGPT as well as the Darcula phishing kit, which help threat actors or help their users generate phishing emails, generate landing pages using artificial intelligence. SpendGPT is unique because it also has a helpful chatbot that is intended to guide criminals through the phishing process to make it easier for them to use the tool in order to make money. And so chatbots and other features are part of the innovative ways that criminals are trying to bake in LLMs in order to differentiate their products from other as a service offerings. And then finally, TJ is going to talk much more about how criminals are incorporating LLMs into the malware itself. But just to preview that a little bit, we've largely seen them incorporating it in ways that enhance specific elements of the malware. So this is not necessarily an end to end improvement, but they are using it to promote defense evasion, for instance, or to automatically conduct discovery in an environment. So these are still largely human operated or human humans in the loop processes. That said, it also hasn't stopped a few threat actors from claiming that they have been able to produce end to end automated malware. And when stuff like this happens, a lot of times, the media really picks up on this and generates a ton of headlines. So we saw this with prompt lock. It gained a lot of headlines as this automated malware tool. But when researchers dig dug into it, they found that not only did the malware not really work as intended, but there had been no evidence that this has actually been used in the wild either. And this brings us to the broader ecosystem that I talked about earlier because it's not just criminals shouting into the void. Their claims get picked up by the media, by researchers, and by vendors in the attention economy. And this is because criminals are playing on a vulnerability that, in particular, impacts media and researchers. Media and researchers are under a lot of pressure to be the first out with a story. You definitely don't wanna be the last one to be alerting on a threat. And so this is a struggle between keeping people aware without unintentionally spreading fear, uncertainty, and doubt. And I just wanna end with talking about how AI another function of AI in producing this in this attention economy, and that is that AI is a powerful tool for actually creating inauthentic content. It's likely that Zero APT, the group we talked about earlier, actually used AI to help them generate some of the collateral around their false claims, making it a little bit more difficult for researchers to go in to try to validate that. It's getting increasingly difficult to separate truth from fiction when inauthentic content looks so realistic, thanks to generative AI. And this is creating a new cognitive burden not just for analysts, but also for consumers of media out there to be able to parse truth from fiction. And with that, I think it's a great moment to pass it over to TJ who will give you more information about how to understand this for yourself and look past some of these headlines. Awesome. Thank you so much, Megan. There's certainly a lot of noise around AI, especially in the AI malware space right now. Every headline makes it sound like we're on the edge of some fully autonomous unstoppable attack wave. So before we let the headlines drive the narrative here, let's bring the conversation back to a measurable reality. We actually built a framework to help defenders separate the hype from reality, but first, let's level set. Let's talk about malware. Traditional malware is pretty straightforward. It follows a logical logical set of patterns. We use tools to analyze it. We can build detections. And even when hackers try to obfuscate, you can really find the underlying code and really understand its functionality, and that doesn't really change. So we're still able to make detections. Now compare that to what AI malware was promised. It's adaptive, self modifying, it lives off the land, it makes decisions in real time. This kind of sounds dramatic. It's almost like a movie. But at the end of the day, malware has to run, it has to execute, and it wants to spawn processes. It wants to connect to something. It wants to steal things and even encrypt. So when it does that, it leaves artifacts. The execution layer is still observable, meaning that we can hunt and detect it. But all AI malware is not the same, and we've defined it in five different categories increasing in cost and complexity. So you have LLM translated malware, is really an operator productivity game where they use, translation for languages to enhance their social engineering operations. Then you have LLM generated. This is where the vibe coding comes into play. Code assistance, faster development, cleaner payloads, it really lowers the barrier for malware operators. Then there's LLM deployed malware. So this is malware that's delivered through AI ecosystems. Model hubs, extensions, poisoned packages, more people are using LLMs, and this is a great attack surface for hackers. LLM driven malware is a runtime orchestration. So this is where the LLM is calling its its model and getting commands on the fly and adapting its behavior. This is where we start to see some of this novelty. And then finally, we have LLM embedded. This is a local model package inside of Halo. This is the fully offline AI capability that everyone dreams about, but we haven't seen really any public samples for this yet. So understanding what category most AI malware currently falls in helps you determine what type of defenses that you can put in place to protect your organization. So to cut through some of this noise, we introduced AIM three, which is the AI malware maturity model. And there are five different levels to help really determine how mature a particular malware sample is, really helping you angle your defenses accordingly. So the first level is experimenting. These are where you're gonna see the proof of concepts. These are attackers testing LLMs, generating code snippets, and it really has a limited real world impact. Next, you'll see adopting. This is where AI starts entering the attacker workflows. This is where they're using it to help write malware or maintain their infrastructure. This is a impact on speed and efficiency. This is not new tradecraft. Then you get into optimizing, level three, where they're intentionally integrating AI into their workflows. So AI improves the orchestration and the Asian techniques that are being used, and operators get cleaner and more refined campaigns, but humans are still in the control here. Then we evolved to transforming. This is where AI starts influencing the decisions dynamically. This is where we start seeing AI driving the actual attacks, and we see this shift to begin with a livid we are starting to see the shift to begin with limited activity in the space here. And then lastly, it's scaling. This is high volume, fully autonomous, AI driven campaigns where we could see a material change in the landscape, and I don't think we're here yet. AIM three gives defenders a common language to separate the capability from the hype and really marking where they're seeing their current threat model. So we've actually been tracking a lot of the AI malware claims, and we published a blog previously on some of these. You'll notice something that's pretty important here. Most examples are clustering around levels one through three. And since our blog post, there has been a couple new instances, void link and honest cube, and those both remain in that in that grouping. So we do see experimentation and orchestration and some LLM driven behavior, but we're not at a fully autonomous AI run campaign scale yet. Humans are still firmly in the loop here. Most observed samples sit in aim three levels one and three one through three. So we're knowing that AI is primarily starting to be used as code writing, which we see, translating content, increased speed and volume, and traditional controls such as identity protection, EDR, network telemetry, these still detect a bulk of the activity. AI is being used as a force multiplier more so than, just completely replacing things for attackers. So why isn't AI malware further matured? Well, I think that there's four major blockers here. The current state is good enough. Ransomware already works. Phishing already works. Why introduce complexity? Also, cost of adoption. LM inferencing at scale is really expensive. We know from some of the headlines of the cost for some of the model training. API usage is also traceable. Theorems and models, and models hallucinate, reliability matters, and introducing AI will just introduce opportunities for those things to happen. And then also security debt and the AI supply chain. AI ecosystems themselves introduce detection and risk exposure for attackers. Criminals optimize for return on investment. If a static tool works, they're not gonna over engineer it because they're gonna be able to get their get their their their operational objectives completed. Now on the flip side, what could introduce some of this AI malware maturity? Well, as defense automation becomes faster and more autonomous, I think that that will drive some of this. LM costs drop day in, day out, so that will that will increase their ability to be able to leverage some of these more frontier models. Standard mature standard maturity of tools will become normalized, so they'll be able to have tools like OpenClaw and some of the new tools that came out and leverage those a little bit easier. And then AI first tooling will become ubiquitous across the the the industry. So, therefore, they're gonna have to be able to speak the same language and operate in a similar manner. If defenders scale AI driven analysis, attackers will have to respond with adaptive capabilities. So offense and defense coevolve. We can parallel this with some of the smartphone adoption as a good analogy here. So in the beginning, you saw in pre twenty tens, smartphones were kind of a niche and early adopters. You were the cool person if you had a really cool smartphone before everyone else had one. Then we got into the twenty tens, twenty fourteens, and then the friction dropped. The costs started to drop, so you started seeing more and more people starting to get phones. And then eventually, in 2015, it became part of infrastructure. If you don't have a smartphone, you won't be able to operate in normal society. We're just coming out of the pre 2010 phase for AI malware and jumping into the friction dropping and the cost dropping here. So right now, malware is AI malware malware is good enough, but AI malware is getting better. The technical friction is starting to drop, and the cost barriers, although they still exist, they're getting a little bit better. We're not at the stage where AI malware is infrastructure yet, but the adoption curve is starting to become predictable here. Now talking back to the paradox that we sit at here. An AI SOC or AI defenses promise, you know, headless alert triage, faster remediation, more coverage at a cheaper cost with AI driven threat hunting increasing the depth of detection. I think all of these defensive scales will require attackers to become faster, stealthier, more targeted, and broader in scope. The metro the maturity of AI malware is going to be likely defined by the defensive maturity that we implement in our ecosystems and not exactly attacker initiated here. So what can we do to defend against this today? Well, we can operationalize our we can operationalize our tools for AI malware right now. So there is a tool in the recorded future called malware intelligence that would allow you to go and search through our malware corpus and identify potential AI malware. There's also AI recording that we can do in the platform to be able to give you an up to date understanding of the trends and the different, the different trends and references to AI malware so that you can apply that to the e n three model. So tracking LLM references and mapping them to TTPs, detecting AI malware in the wild and understanding the current maturity of that landscape, monitoring suspicious model behavior, and and basically making periodic attempts to to understand the current maturity level of the ecosystem so that you can adapt your defenses accordingly are gonna be important factors here. So let's step into a example here. If we look at LameHug, which is one of the first purported AI malware, we can actually go and do a practical threat hunt in malware intelligence just looking for samples that are reaching out to LLM APIs. So this one is find me the samples with DNS request to router Hugging Face, which is one of the LLM APIs. This identifies lame hug leveraging AI the AI malware infrastructure. And the key here is that the AI malware has to interact with some of these ecosystems in order to have the LLM advantage. So that interaction is something that we can track and monitor to identify opportunities for detection here. So when you look into that, specifically here, you're able to see some of the some of the executed commands. You could see drop files. You can see process calls and where it actually touches the disk. It's interesting that defenses are not really changing. They're just requiring you to be faster. And you can also apply other other types of detections here. So you can monitor the different types of references to AI malware. You can categorize them to aim three, and then look for different things in that malware to build build detection, such as knowing the families by just extracting the signatures for those from malware intelligence or insect hunting packages. You can look for static LLM characteristics, so looking for LLM formatted contact content like, prompt or embedded model files, so files that would be associated with that. Lastly, you can look for calls for LLM libraries on your in your environment to identify potential malicious AI usage by malware. These things leave footprints, which allow you to track them, and then you can measure them with aim three and then prioritize your defenses. And with that, I'll get back to Megan to talk about the key takeaways. Great. Thank you, TJ. So, again, just a reminder that if you have any questions, please pop them into the q and a box, and we will be happy to address them. But I wanted to wrap up with a couple of take home notes. So we've seen that AI currently acts as a force multiplier, and the more skilled the threat actor is, the better able they are currently able to utilize AI in order to improve their operation. But as TJ mentioned, we're not yet seeing full end to end automation being carried out by malware. And so he also gave us some helpful markers to look for as a potential way to consider where this technology is going. We've also seen that strategic hype around the use of AI gives cybercriminals an increasing competitive edge when they're advertising different as a service products as well as increasing fear, uncertainty, and doubt for cyber defenders and the general public. So it's certainly a deliberate tactic that they're using in order to enhance their operations. And then finally, frameworks like a AIM three can help defenders understand and importantly communicate what the actual threats are posed by AI enabled malware in order to help you better defend your networks. So with that, we can take on some of the questions that we've gotten in the q and a. So I'm gonna start with one question from how much of cybercrime is state sponsored, especially indirectly, to maintain plausible deniability? So this is a great question, and I know I've focused a lot on cybercrime, but we know that there is some overlap between cybercriminals and state sponsored actors. So it's difficult to know exactly. Recorded Future has actually put out a series of reports that really do the work in tracing individuals between, in this case, Russian intelligence and security services and Russian cybercriminal groups. So I definitely encourage you to check out the dark covenant reports for more details on Russia specifically. But if we take a step back more broadly, we've seen state sponsored actors work indirectly either with tacit agreements that if as long as criminals don't target assets the state finds important, they can go after their enemies with as much impunity as they want. And we've also seen some more direct engagement to the point where state sponsored actors are actually carrying out cyber destructive cybercrimes pretending to be ransomware. So we've seen some Iranian threat actors use this in this capacity. And this has to do with hype because a lot of state sponsored actors are also trying to benefit from this fear, uncertainty, and doubt around different types of cybercrime. So hope that was could do yeah. and I could do the next question. So how is the EIM free model tied to the tape of AI malware? So those are two different angles to look at it. So first, you wanna make sure you bucket the different types of AI malware, what it's actually leveraging because that's gonna determine how you're actually gonna incorporate defenses. So from a a the LLM translated standpoint, you're probably going to want to do phishing exercises so that people understand the nuances between language, just to understand when someone has taken content and just translated it into a native language for their victim, and then they're passing those along. But then when you come into something like LLM embedded, that's when you're gonna start looking and incorporating reverse engineering techniques to identify those models that are embedded inside those malware payloads. So it really the the type of AI malware determines how you would impact the defenses, and then the maturity gives you a sense of urgency of how important you're going to be able to, put those defenses in your environment. So you may not wanna have to inspect every single malicious payload for, MLMs embedded at this point. It may be more important to to leverage the phishing deterrence or resilience training. So that that would be how you'd you'd be appropriately applying those two different models. Great. So the next question. Do you say that there are three main malware models that are verified? So this goes back to what I was talking about with the criminal large language model. So, essentially, LLMs with no safety guardrails that are specifically designed to help criminals carry out their operations. So when I mentioned the three models, these were the three that we detected in 2025, and you can read more about them in our state of security report. We talk about what those models are and specifically give a little bit more information about them. But those aren't the only three models out there. I think the last time I checked, there's maybe eight or 10 active models that we are tracking. There may be a bit more. And the reason why I'm a little uncertain about this is that these models tend to rise and fall with the developer's interest in maintaining them. Sometimes other if they're especially if they're using open source tools, sometimes those can get cut off. Or if they're using commercial tools, they're detected and then cut off, and so they're unable to, keep their jailbroken tool up. So, there's definitely a lot of flux in this environment that, that we were seeing in tracking models. But, yes, the three main the models that I referred to are from, emerged in 2025. So and then another quick question is to share a slide with team. We're actually gonna provide these slides to you guys in PDF format after the after the webinar, so you should just check your inbox, and you should be able to get that shortly. So let's jump to the other question is, what is the best documentation link for malware intelligence in the threat module? So, actually, we have a white paper that if you go and look, search through the support documents, you should be able to find a, intelligence package white paper that actually shows some of the best queries to do in malware intelligence, and it does include, hunting for AI malware in that. So that would be one of the best ways to really get started in malware intelligence. We actually use a lot of those queries because we provided it in that document, and that's how we find them oftentimes. So that's a great great question there. K. And so here we have a question that, TJ, I think that both there's probably gonna be a two part answer to this, so I'll tackle the first part. A number of vendors are making comments that AI is a factor in x percent of attacks with x being a high number, and these are coming from reputable firms. So is this such a sales ploy, or is there truth to it? So as TJ mentioned, it is possible to kind of track how often AI is being used in attacks. And especially when these reports are coming from some of the AI researchers themselves, they're actually able to see how accounts that are linked to threat actors are actually using the malware to enhance their operations. So I think maybe the parsing is not necessarily in the exact number of attacks, but really what do we mean by factor. So you wanna look at how these threat actors are using them. And in a lot of cases, it's not too dissimilar from how we use AI currently to enhance our legitimate operations. Right? So threat actors are using them to write code, to produce content like phishing emails, you know, the corollary being maybe in advertising or sending out marketing emails. Like, we're using AI as a tool and so are the threat actors. And so this is not necessarily too surprising the threat that vendors might be claiming this. So where you might see things being a bit more of a sales ploy or things being a bit overhyped is when they are trying to, unfortunately, increase some of that fear, uncertainty, and doubt around that the only solution is to either buy their product to actually fix this. So kind of looking into what exactly they're talking about when they're putting that statistic and what is the intended action that they want you to take. Yeah. Absolutely. I think AI is arguably a factor in a lot of different, work operations and in our personal lives at this point. So I think it's really important, and this is why we have the categorization of how the LLM is enabling the different payloads is how it's being impacted. So if they're using it for cogeneration, that's gonna be very different from that orchestration mechanism where it's LLM driven. And really understanding the differences between those allow you to sort of wrap your head around how important it is to apply certain defenses in your environment. And I think there's another one last question here. Do your findings on the use of AI in the malware ecosystem also apply to vulnerability discovery? So, that's a that's a broader question here. I think the the framework for AIM three is really focused around AI malware mainly because it is a microcosm of the larger AI, the the larger AI paradigm that we're going through currently. Vulnerability is a subset of that. The the hard part about vulnerability is that, we do and as we do see, that is actually one of the most, used and on the forefront of a lot of the the AI maturity front. We're seeing, you know, nation state actors and and, prepositioning networks being leveraging LLMs in order to increase the velocity and scale of their operations. And the interesting part about is that is that when you're actually doing the defenses or you're doing the analysis, it doesn't look any different from what a traditional attack would be except that it's faster and more articulated, and it just looks more mature. So it's hard for us to really measure that unless we have some of the visibility of a frontier model that would be able to actually see the prompts and see some of the interactions with LLMs in order to sort of understand the modus operandi of the attacker. K. So we have, a question that, essentially boils down to what are the main skills and tools to learn as security specialists in order to kind of address how threat actors are evolving. And so, again, I think we can take this maybe in two parts, kind of thinking about both how to cut through some of the FUD as well as how to kind of look for some of the more tangible indicators and as well as adapt your security operations center for some of the real world impacts of AI. So thinking about some of the FUD, and especially because, as I mentioned earlier, AI is going to actually make it a lot easier for threat actors to produce inauthentic content, to produce misleading content. So using synthetic data, for example, to try to convince you that they have actually stolen data from your network instead of, you know, just claiming to have done that. Now this is something that threat actors have done in the past to try to essentially trick targets into thinking that they are under cyber attack. But with generative AI, they're going to be able to make much more convincing collateral that's going to be much harder to parse. And so I think being able to get to some of the ground truth, finding evidence on your own side in order to piece together what actually happened in attack before responding a bit more emotionally to a high pressure situation that threat actors are trying to create is going to be an essential tool. So, essentially, making your security team psychologically resilient to some of these manipulation tactics in addition to the technical resilience that you're trying to build? Awesome. And then I think we have a couple technical questions, and I'll I'll blast you real quick here. So the first question is, is there any unique signature to identify whether malware is created using AI or non AI? The current patterns for understanding whether a or in where you see a lot of researchers are saying this is AI AI generated is mainly some tells. So I I know the m dash is, has is going down in infamy as a tell for potential AI generated content. There's similar things in programming that we can identify. So certain patterns of of code, certain, like, documentation. So, like, if they're using docstrings a certain way, you can sort of tell by the the language in that if it's decompilable or if it's actually, using certain, like, obscure libraries that people usually don't use. You're gonna have a really good tell of whether it's l m generated or not. And so that's the current state. This is always evolving. I remember, with Python specifically, there was, docstrings where you would actually be able to see docstrings, like documentation strings in the code, and you'd sort of see that there it was a one generated. And now that's kind of evolving as those frontier models start getting better and better and more humanistic. The next question here is, is Hugging Face code bad at all? Are you recommending blocking it? No. This is an important distinction. So the usage of AI is actually a good thing for some customers, but by malware, it's very bad. So just like every actor, they will leverage live off the land, leverage things that, you know, could be seemingly benign in good hands, but can be malicious in others, they're gonna do the same thing. So this is just a tool just like any other for a malicious actor. And so understanding what the provenance of that that interaction with Hugging Face is important. If it's a developer developing something for your company, you're obviously gonna lean towards that's probably good. If it's a obscure or anomalous usage for a user that probably doesn't know what LLMs are, that's gonna be something alarming. Hey. Another question. Do you think CrowdStrike and other vendors naming criminal groups is adding to the problem? So I'm this is a really interesting question because I think this is absolutely a double edged sword. So this is definitely a very necessary naming threat actors is extremely necessary from a cyber defense standpoint. Right? Because how do we know what we're talking about if we don't have some sort of threat actor group name? And every different vendor has a different methodology that they use in order to assign names, assign threat actor groups. And, yes, some of them come with pretty cool looking branding, but I think that that's something that it's hard to avoid if you want to have a rigorous cyber threat intelligence process that you know, where you can actually track threat actors. That being said, threat actors also know that they're being named. They're kind of they're tracking this as well. And so that's exactly what the point that came up with the group scattered lapses hunters. Right? They know that all of these names have brand recognition, and so using the three of them together helps amplify their group's notoriety. And so this is another way that criminals are kind of manipulating how defenders are looking at them, how researchers are looking at them. I mean, criminals are kind of trolls, so they definitely are trying to manipulate some of our processes that we're trying to do to defend against them. So, unfortunately, I think that, you know, we're not gonna ever be able to fully avoid kind of adding names to some of these groups. Awesome. And then here's a a deeper, but really good question. So as attackers evolve and their techniques and malware becomes more sophisticated, What are some key points and challenges corporations and enterprises need to address in order to be procrast proactive enough? What are the main skills and tools to learn as security specialists to address that, especially when more AI powered solutions are beginning to be deployed nowadays and the malware security blue blueprint is not yet all that mature. And there's more and more security researchers getting into this field. Yeah. This is a long question, but I think it's a really great one. I'll start, and then I know Megan probably has some stuff to follow-up on. Really, it's understanding the fundamentals and the basics. As I sort of pointed out throughout this this conversation here is that AI malware is similar to malware, and we can actually detect it still in the same mechanisms that we do regular malware because, ultimately, it's doing the same thing. The game hasn't changed yet. And so being really good at your defenses is gonna be important because what AI has changed is making this a lot easier, lower barrier to entry, and then also scaling is gonna be higher, and speed is gonna be, you know, lightning fast. So making sure your defenses are, able to be resilient to those those efficiencies are gonna be really, really important. So understanding how AI is used, understanding where your threat surface looks from, from afar, and really making sure that you're tailing your defenses to cover the things that are your risk profile is most vulnerable is gonna be really important for you. Megan? Yeah. Absolutely. I don't know if I have too much more to add to that, but just to consider, again, comparing how we're seeing it today and then also trying to keep an eye on the future as well. So kind of using that maturity model to help assess where are the most likely advances that threat actors are going to be able to take. And one comparison that I like to make is that if we can do it, they can do it. So as we are able as we're able to use AI for more advanced coding, as we're able to use AI for more complicated prod projects, we should assume that criminals can do the same thing. Another question here is, we've avoided talking about anthropics report. What is your take on their l four claim? Well, actually, in the blog post, which I think we'll send out, it's it's on the Recorded Future website where we talk about AI high burst reality. We actually do address that. I particularly don't think that it's a fully fully autonomous capability. It's reaching towards it, and it is the closest to l four that I we've seen. However, because there are humans in the loop, I think that it was important to really just make sure we have that distinction. And with the Anthropix visibility of that attack, which they they're obviously gonna have large visibility as a frontier model provider for that, You could sort of tell when they made that distinction, they said that there were humans in the loop, which was a key differentiator in that l four designation. Right. And, basically, the more sophisticated a threat actor is, the more sophisticated applications are going to be able to find for AI. So they'll be able to simply use the tool much more effectively. Great. And then I think the last question that we'll take, what does FUD stand for? Again, that is fear, uncertainty, and doubt. So shorthand for kind of creating confusion tactically. Great. Well, these have been great questions. Thank you so much for sending them in and for your attention this morning. We can also wrap up in a follow-up blog post as well to if there are anything that we might have missed. And, of course, we will send out the recording to all participants following this webinar. So thank you very much for joining, and we hope to see you at a future webinar soon.