Selena Larson, ProofPoint | theCUBE + NYSE Wired - ProofPoint Protect 2026

Palo Alto Studio Connection Silicon Valley and Wall Street. I'm John Fost here with Dave Volante, my co-host, host of the Cube. We are here in San Diego for Proof Point Protect 2026, part of our NYC wired program and our open community. We're unpacking all the data around security in the enterprise. As ages and data become more prevalent, cyber security posture changes. We have a cube alumni back on threat research. Selena Larson back formerly on at another conference Mise which is no longer around which is really one of the premier threat conferences really talking about the bad guys on all the nation states. Great to see you. >> Yeah thanks so much podcaster researcher. So first of all congratulations on uh the continued podcast and getting the content out there. >> Thank you. >> But I really want to dig into a last change since in the past few years. Last time you were on was 2023. Mhm. >> AI continues to thunder away, but the game has gotten much higher level action >> on the threats, more organized, more complicated, and more effective. >> But AI has come on the good side, too. So, what's different? Give us that lowdown between, you know, 2023, the old way seems like yesterday and today. >> It does feel like yesterday, actually. I think about I was like, wow, that was three years ago. Um, and that really was I feel like in 2023 when AI was just emerging, people were using it uh in their own environments and threat actors had started to adopt things. So, I would push back a little bit on this idea that things have gotten better and attackers have gotten smarter. I think that in some ways they have, but in some ways they're regressing. And I think it's been really fun to see because for many of the threat actors, they're using AI in the same way that defenders are to streamline their operations, to um make their social engineering a lot better, right? And some people are using it for marketing or styling things like that. And uh they're also using it in their attack chains. But what's really interesting is the good threat actors are using it very effectively and are cutting down the time it takes to develop something such as malware or um spam delivery, things like that. But for bad threat actors, the sort of bottom of the barrel folks, the low hanging low hanging fruit so so to speak, they're adopting it. And yes, it does make some of their attacks a little bit more believable. their social engineering is a lot better. You know, we were talking before this started. The emails look really believable. The the um the language is there. The design is is is pretty good. But what we've seen is actually regression in techniques. So, they're not necessarily as effective. So, they'll leave things exposed on the internet. >> They get sloppy. >> They get sloppy. It's very, very sloppy. >> All right. So, you're pushing back on the fact that um AI the good guys are not getting better. You think they are getting better. I think the good guys are getting better >> and the bad guys are bifurcating into the elite. >> Yes. >> And the idiots. >> Absolutely. Yes. >> Well, I I say this because um um a quote on the cube recently, I won't say who the person was because it may be taken the wrong way. Uh she said, "AI makes smart people smarter and dumb people dumber." >> And I'm like, "Okay, that kind of hangs." But let's just get about the implications. It's kind of true. If you really don't know what you're doing, you get false sense of security. And you were referencing that because I think that speaks to some of the sloppiness. >> We used to call it kitty scripts back in the day. Remember those those kind of porative terms, but what has that done to the to the artifact? So what's been the impact of this? Has there been a backlash in amongst the bad guys like hey say let the let the big players take the big game boards. It's just too much for everybody. >> You know that's really interesting that you bring that up. So Wired a couple of months ago published a great look at how threat actors and hackers are responding in some of the forums to the adoption of AI in tooling and and and marketing and things like that and they don't love it. And so a lot of this >> don't love it in which way >> they well they were like get this AI slop out of out of >> they have their own slot problem. >> Well yeah cuz they're like I don't want to read this like can you I just want to buy this tool. Why are you adding these things that I don't need? It doesn't work anymore. So they are getting frustrated with the thread actor adoption of swap and at the same time you have thread actors that are using AI tools to build things like fishing kits for example using um device code fishing right so they'll make a really sleek landing page and says oh this you know impersonating Microsoft or docuign and then they'll have a code to input it to to legitimately associate an application a malicious application with your Microsoft account but if you look at the rest of the attack chain the panel itself will be exposed um the domain won't be very believable. The lure itself, if it's using something like social engineering, sometimes it might be pretty good. And sometimes the email is just blank because they don't know how to actually do like mouse. >> They're bad at marketing. >> Yeah. They're just they're just not very good at marketing their own security. >> Yeah. Yeah. So, we have so we have these sort of these these slop actors that are doing a lot of that. But then you also have some pretty creative thread actors and and thread actors who have done malware development in the past who have created really unique and interesting loaders or botn nets and things like that. they're also adopting AI and you can see it in their code, right? Because nobody comments their code like an AI comments code. Um, and so you'll you'll see that throughout and they'll say this is what this JavaScript does or this is what um where you can like insert a specific Python or this is a a specific command or PowerShell um or even if you're looking at um something like um something like a loader, it'll be like insert the file here, like the file name here, things like that. So you template >> it's a template. Yeah, it's very much templated. and for the threat actors who are really good at it. It can be fairly effective. Um, and you can and it's just sort of making them better. But I think that's that's fewer than what we're seeing with the adoption of slop. And I don't want to say that just because it's sloppy doesn't necessarily mean it's effective because we've throughout history have seen social engineering that might not be super sophisticated be very effective. >> I mean, if someone's taking a one shot at something, they don't really care what it looks like after. It's kind of like the software factory cares about >> garnage whatever bad code laying around. What has it done for me? Let me ask you this on your research. There's been themes of hey the the basic blocking and tackling kind of techniques are just getting better and faster. That's kind of a proof. Got to deal with those with machine speed. But there are new things emerging. What is new? What's emerging that you haven't seen before that AI has enabled? Has it been better coordination? has it been just new ways scale understanding better social engineering combined with other things. >> Yeah. So I think scale is a really great one. So historically what we've seen are thread actors will tend to target specific languages, specific geographies and with with one campaign and then maybe a few days later if they are a threat actor that targets multiple then they'll switch to a different language. What we've seen in some cases is up to like 15 or 16 languages and lure themes that target a broad variety of geographies. It's a really uh carbon copies of each other, but the language changes. And so you can tell that they're using some type of automation, some type of scripting to be able to clone essentially the original um the original lure. But because AI is really good with natural language, they can make it seem a little bit better. Whereas historically, if you were trying to clone something into 16 different languages, it probably wouldn't look natural. It wouldn't look legitimate. So things like uh for example, Japanese or Korean, uh the the AI has improved those templates quite a bit. Um and yeah, so we'll so we'll see a higher volume in campaigns. Um the other thing I think that we've seen more of is just experimentation on filtering. So what I'm what I mean by that is when a thread actor is trying to send high volumes of of URLs for example and they're trying to make sure that only you get it there is a trick that they'll incorporate either something like geo filtering um IP filtering or something like a capture where you have to solve a puzzle in order to get the payload. They want to make sure that it's a real human being that's actually getting the fishing payload and not me and my sandbox getting that. So we've seen a lot of experimentation in that aspect of it. >> So they figured out the honeypotss, they figured out all the little tricks, the sandboxing. >> Yeah. So there So they they kind of know how to do some some sandbox bypassing. So it'll be like, "All right, I'm going to come up with this capture that no one's ever seen before. Some sort of weird puzzle or >> No bicycles. >> No bic. Yeah. No more bicycles. No more stop lightss. No more Yeah. cars. How many cars is this?" Um, but I I think I've seen ones that are like little like fish emoji and then it's like figure out the strawberries like weird just weird things like that. So they've they've they've changed that up a little bit and with every new iteration of that it's on us as defenders to be able to come up um with more sort of dynamic detections. And so one thing I think is really cool from our perspective as the defenders is that um historically you know static detection on IoC's and landing pages and stuff could be pretty effective because they would be used for a prolonged period of time. But now the time of the or the length of time that that static detections are valuable or useful has really decreased because of the rapid iteration and development of the the payloads and the attack chains. So it's on us to be really creative. It's like, okay, how can we come up with some really interesting dynamic detections that we can anticipate? Okay, we think that they're going to try another type of filtering so we should create detections for this to to preemptively block it. Essentially, >> the intelligence, that's where I think to your point about the AI is getting better on the good side is because, you know, there's real time stuff on the defense. Okay, we know what that is. Kill it right away. Um, but then there's like the more complex things that look normal like >> that's the mechanisms are all working. There's no malware. Email comes in. Oh, you run accounts payable. >> Yeah. >> Oh, there's all the suppliers on your email. Let's just email them, too. One owes money. That's So, there's all this, you know, kind of like living off the land by accident on purpose. You're like, okay, I'm not living on the land like hanging around. I'm actually just discovered >> all these new people. >> All I got to do is get one of them. Yeah. >> So, there's all these techniques, but that doesn't look like it's def like breaking anything. So the psychology of the data was why I'm intrigued by this knowledge graph element because you might say oh John is running accounts payable he's on vacation and in Slack he already told the people there's some data so like this other context >> how is how is that defense helping any new ways to pull in that the psychop side of it >> to the defense side >> yeah so I think that's one way where AI can be really interesting in terms of behavioral detections. So, like you mentioned, oh, so and so is out of town. Um, also things like misdirected email, if there is a CC, a BCC or a reply to that's unusual or outside of the scope of like what someone might typically reply to. Um, then you can say, "Oh, wait. This seems out of the ordinary. This seems like, you know, that something that we should flag on. Are you sure you want to send this email to this individual?" um or if you receive it on the recipient side something like this is a this is a suspected you know or a domain that doesn't have a longevity this is this is suspicious so I think um things like that can be really effective um and then same thing with like you know like looking at data loss and the data loss landscape and the insider risk too right because when you're a threat actor who has done an account takeover and gained access to some something uh I uh I think detection opportunities are something like malicious apps that might be added to um to the uh account itself. So this is something that's suspicious. This is out of the ordinary. This does this name doesn't align with our existing enterprise applications. This is something that we should investigate. Um and then also to downloading uh various uh files, trying to excfiltrate that that type of data. So kind of paying attention to where it's going um and what authorization and what uh sort of identities are enabled within accounts. My uh my my interest on agents is interesting where because agents remind me of when Facebook came out with misinformation. It wasn't the network effect and the the graph. It was the content was the payload. >> And Facebook just wasn't paying attention. Yeah. >> So the theme we're hearing now in AI is that we want to pay attention to the trust. >> Yes. >> The agents will do what they're told. >> Yes. >> So how is agents because they can be weaponized >> just like Facebook was weaponized for misinformation. That conversation is starting to lurk around in the mainstream now. We're saying, okay, agents are going rogue, but are they really going rogue and also if they can get weaponized taken over, >> that's an interesting dynamic. What data are you seeing on the agent front? Because this is then brings in the question of okay, I got the good agents and the bad agents. There's espionage going on. So there's like they're crafting their own civilizations. >> Well, that's an interesting way of putting it. Um I think >> can they form can agents form as a digital twin of the threat actors? >> Um I think if they are programmed to do so. So I don't necessarily think that we are going to have some sort of rogue um rogue robots like help open the pod bay door how no um I don't think that that's necessarily coming up in our in the near future. Are there any threat land uh vectors around agents that kind of give signaling? >> Yeah. So, one of the Yeah. One of the the more interesting things I think about about this and is is supply chain for example. So, um and I actually just brief aside when you mentioned you think about it with Facebook a lot as sort of the original agent. When I first got Facebook, I remember it was posting everything that I listen to on Spotify to Facebook. And I was like, "Wait a second. Why is everything that I'm listening to on Spotify automatically being posted to my Facebook page and people are commenting on it? This is so weird." And that was like, "Wait a second. I don't have the proper restrictions in place. I just sort of added something to my account. >> Install on some appreh." >> Yeah. And I didn't necessarily understand how it worked or what it was talking to or what data was going to be shared with what account was going to be shared with. >> At least she didn't throw MySpace widget on there. >> Oh yeah. Wow, >> those are the first generation of spam takeover of Facebook. >> Yeah, I mean that's how I learned to code was MySpace that marquee scroll. I got that. So, but but that's a really great analogy that you bring up in thinking about Facebook because that's how I think about agents. If we're just installing any sort of agent on our devices without the governance aspect of it and without understanding, you know, what accounts are talking to it, what information is being uploaded, what is this AI doing with this information? Um, who has access to this information? And I think I think it's a little bit uh parallel draw some parallels with the cloud as well too, right? Because it used to be like we're just going to upload everything to the cloud and then guess what that was >> bucket was open. The S3 bucket told me about Yep. How many data breaches did he have because of exposed S3 buckets? >> And that's the same thing no one told me. So there's kind of a rules of of the road. So I guess my question is as you look at as you look at the threats >> um that's just humans got to keep up. So it's kind of like basic blocking and tackling but also that's net new. Oh, we didn't know that. >> Yeah. Well, so I think you know going back to this idea of supply chain and that I think is something that's pretty interesting because if you look at um essentially maybe agents pushing malicious commits into a GitHub for example and then having a uh uh the autonomous development pipeline just automatically accept things without review that could potentially lead a malicious agent to upload itself essentially into your own network. So if you are just relying on agents to uh do this this this security and do the auditing and you're just pushing accept on things um because you you just trust that it's going to be secure. I think that that's where we're going to get into some issues and we've seen poison like supply chain poisoning in Python packages for example or um on GitHub or in um various coding repositories. And right now what we're seeing within the enterprise is an automated continuous integration and a continuous development pipeline. So organizations are just relying on things that are already in open source or already in um these various gits. >> So the CI/CD pipeline was never really programmed to do software supply chain validation except for containers. >> Yes. Right. It was it was something that we're just like okay we're just >> it's working until it doesn't work. >> Yeah. Yeah. Right. Right. >> Threats here. >> Yes. Exactly. We're doing a sprint. We have this and oh wait a second now we have potentially rogue AI. >> Yeah guys I got to get my margarita in 10 minutes go. Like >> exactly >> that's human error. That's back to the human piece. That's why I like the proof points perspective. It seems to be a much deeper >> thought process around >> the endpoint relationship to the user >> and then the agents on the inside. >> Well, >> so there's a whole another level of two-sided >> Yeah. >> human agent relationship. >> Yeah. Well, and I think one of the great things that I love about Proof Point and have we've really adopted this since I since I have embraced it since I've been here is this human centered security and I think that you know as we're moving to an agent centered security there's always going to be the need for the human in the loop and I think that AI and and agents and agentic development and agentic socks for example um like security operations that are AI enabled what we're doing is effectively just giving the human more power more tools more resources human. Don't you think like a manager would have direct reports that humans should be responsible for agents because makes sense like hey what happened I don't know an agent got loose I don't know who's from fleet it's from reports to that's Selena's agent not my agent proven >> don't point fingers here that wasn't my agent like not my agent his so I mean but this comes back down to okay whose organization because it's work it's an extension of work >> almost look as you're an operator of agents a manager of agents the human has to take responsibility. >> There does have to be some responsibility because I mean, fundamentally, our AI is only going to do what people tell it to do. >> And you know, we don't have AGI right now. We don't have robots that are like, you know, building themselves. And we're essentially they're all reliant on the person that's telling them what to do. And we do run into some issues where if you have someone who isn't necessarily trained or knowledgeable about what they're doing, for example, a threat actor creating malware but not knowing how to put their botnet panel behind, you know, secure resources, so you can very easily take it down. Um, you have the same thing from a defender perspective. So if you're not, >> so it can go, it can get loose without it's just by accident. I was trying to do this, but everything else happened. >> Yeah. So that's why I do think that there are some like really great opportunities for um for automated governance, automated um detection of some of these rogue threats and making sure that you're you know you have these um tools and protocols in place that your your AI can tell the person are you sure you want to do that or uh you're going to this data is going to be sent here is this compliant with your >> talk about the um the organizations that have you know have become companies I mean mal malware's been out there and ransomware you can buy ransomware as a service. We've covered that years ago. >> Yeah. >> Are there yet agentic as a service acts? >> I'm imagine this just another extension of as a service >> on the higher end. Is what's going on with the big actor, the threat actors? >> Yeah. >> Is there any movement on crushing them and getting >> crushing them? Yes, we're always crushing them. >> Well, Mwise, we talked about not just, you know, >> give slap on the wrist, kneecapping them out completely and arresting them. And that was something that came up >> like taking them off the streets literally. >> Yes. So there is a lot of efforts and continued efforts on that. Um for example, one of the efforts that we participated in recently was the disruption of the tycoon uh fishing kit. So it was a multiffactor authentication of fishing as a service and uh we were able to work together with some private public partners to be able to um do that disruption. a civil lawsuit filed by Microsoft in uh in with support of international law enforcement was both a online disruption and it did have some real world impact with you know congratulations taking people down. Yeah, I was very excited. >> Get a challenge coin. >> Uh I didn't get a challenge coin. The challenge coin is the friends we made along the way. >> Yeah. Yeah. Know but it's also nice because that's like you can actually that's tangible impact. >> Yes. Absolutely. >> Not just reconstituting with the gang or whatever. >> Yeah. Well and we're seeing that now with um various malware. So you're asking if there's AI as a service. What we are seeing are different thread actors using AI to build services. So u for example fishing as a service like I mentioned before device code fishing as a service is something um that is quite popular in all of the device code thread actors are using AI to build these tools. Uh we also see um for example um clickfix as a service. I'm not sure if you're familiar with clickfix but essentially clickfix is a technique that uh thread actors will use to compromise. They'll compromise a website and then you'll see a little popup that says your uh Windows is out of date. Copy and paste this com like Windows run which opens PowerShell and then copy and paste this command and then it'll be it'll be up to date. Of course, that's copy pasting and run PowerShell that installs malware and etc etc. It's bad. But that was quite popular with a handful of pretty bespoke and and sophisticated thread actors. But now what we're seeing is a threat actor cluster um that's actually doing that as a service. So thread actors can pay them to be able to use their AI generated clickfixes essentially. Um so they'll you know be be compromising these websites using this this actor created clickfix. So they'll they they are using this automation and tooling that a thread actor is selling. So I'm not sure if we've seen sort of AI agentic solutions in the same way that we're seeing AI enabled threats be sold and repackaged as as a service. >> You know this is so fascinating. There's so much going on and love the success of the take down of the organization with the law enforcement and taking them offline with disruption. Um what I'm noticing is is that there's been a real generational shift. We're seeing it at the political level. Yes. Of I just don't trust anyone anymore. I don't care which party you're from. Seeing that now play out on the on the national scale. There's kind of a CEK version going on. >> And I'm an ex generation or so. You know, we're the last feral generation as has been been described, but you know, I'm old compared to the young guns coming in now. But like there's a whole mindset of like there's a whole new way. Are people changing their expectations on how they interface? Because most people that didn't grow up on the internet I know what a domain name looks like so I can tell by the URL but techniques are so good now that the younger generation are seeing new popups. >> Is there a trend to go headless and go you know I won't say analog but like minimalistic. Yeah. >> On interface. Are you seeing any generational patterns around the consumption side? that's more safer. >> Yeah, >> maybe less functional and generational shift on the attackers. >> That's a great question. You know, I have seen some Tik Toks from the Zoomers and Gen Alpha that are basically like I want to go back to the '9s and like they're they're like they're missing the flip phone that they never experienced um or no one calling them up. >> Well, I mean, I have seen that all over just in my own neighborhood and in I live in Washington DC. There's a big movement to do like a no phone two weeks where you lock up your phone and just kind of >> try January. >> Yeah. Exactly. And like live analog, but I think that there there is some of that because I feel like, you know, Gen X and I'm I'm millennial so we were you know kind of grew up as internet natives and we never trusted really anything. And if you look at historically um looking at the the behaviors and who gets defrauded the most, it's it's usually the older generation. But now you're seeing a lot more young people that are that didn't have that digital native experience that that transition from >> they'll sign up for anything, >> right? Yeah. Yeah. Or click on any link. >> Yeah. Or on TikTok, they'll be like, oh, I you know, would you like to be an am a brand ambassador and here's some free clothes if you send us, you know, 50 bucks or whatever. So for shipping and so you have this sort of like this this generation where they're getting scammed a lot more than I think historically we were able to. And I do think that things like AI, deep fakes, um, improved social engineering really do play a role in that. And so I think the natural progression of that is for people to push back and say, "Wait a second, like I don't like this. I want to go back to when it was the '9s and we had, you know, we had to call me on my house phone to to hang out." Um, so I I do >> at 10:00 on the corner, like, you know, actual real communication. >> Yeah. In fact, I'm I'm mentoring a teenager and he's like, "I actually prefer phone calls. Can we can we have a phone call as opposed to like text or email?" And I said, "Yes, we absolutely can." >> Of course, you're a podcaster. You're like, "I'll talk all day long." You know, like >> I love yapping. >> It's funny. Podcasters love I was talking to someone in support. They said, "Anyone who likes podcasts like to call support the support number." So bloggers would all email. >> They all email. Look, I don't mind waiting on hold. You know, I you know, I have to I do have to say though, if we're talking about people being frustrated about technology, I've seen so many people be so frustrated with like AI support bots and like this the call centers. is just like I miss talking to a human. Like I just want help. Like >> we know you're not real. At least be good, not real. >> Yeah. Just pretend to be good at this. Yeah. Yeah. So that's >> all right. So what's the coolest thing you've worked on the past year that or couple things that were either cool you didn't see coming or cool you knew was coming? What were some of the cool things you worked on? >> Well, I actually am I allowed to say something that I didn't actually work on but my team did and it was really cool. >> It's something cool that people might see and be motivated. >> Yeah, absolutely. So, one thing that we have recently published information on is something called um the blue moon exploit chain. And this was really interesting because my colleagues discovered an attack chain that was used by a variety of Chinese uh espionage thread actors and it chained together multiple vulnerabilities in Chrome and Windows. But what was really interesting was and if we're talking about AI and the use of AI um they uh the the actors actually were able to create something called a patch gap zero day when someone identified a flaw a legitimate researcher identified a flaw in Chromium published that information here's uh here's this bug uh Chromium and the dev tools were able to sort of um fix it and adapt it but there's a two week two to four week gap between when something is patched in Chromium and when it's rolled out broadly. for everyone on Chrome and using the Chromium build. So between the time that that was reported and the time that it would have been patched, the thread actors were able to create um an exploit to uh to to to use this vulnerability. And um while we don't know 100% for sure, we do suspect with fairly high confidence that the thread actors were able to essentially reverse engineer and create with AI this uh exploit that was related to um that this this post um that was actually the bug report for for Chrome. So u my colleagues named it the blue moon exploit chain which was very interesting and it was used by Chinese adversaries but it has but since been adopted um by more threat actors >> so the date reverse engineer the note or the timing the clock's ticking go get it exploited. >> Well it was the actual it was the actual report. So these these bug reports and open source bug reporting is very very like detailed and so you want to be able to say here's everything I found here's how to fix it and they'll say yep thank you and then then go ahead and do that. And while it was patched, it was fixed. Um, it does take a little bit of time to sort of roll that out in the patch gap. So, that was actually pretty interesting and I have to give a shout out to my to my team for for finding that. >> All right, nice prop. Anything else? >> Um, I would say I don't often see interesting AI developed malware, but there are a couple of of recent malware families that we have found that are created by pretty sophisticated malware thread actors that are are making uh making voters a little bit more interesting. So we'll hopefully have more to share on that in the coming weeks. >> We'll we'll wait to see. Get busy working. So thanks for coming on. Thanks for the great conversation and the insights and commentary. Appreciate it. >> Thank you so much for having me. >> John F with the cube. We are here proof point protect 2026. It's part of the NYSC wired program and community. Thanks for watching.