TL;DR: This week's pod is a conversation about AI and "Open-Weight Models", and if you aren't intimately familiar with what these are this is an episode for you. Wes has a lot of insight, and this is a discussion rather than an info dump. Guest: Wes Widner Description The most disruptive AI feature might not be a new chatbot trick; it might be the ability to download the model and run it yourself. Recorded live at B-Sides Atlanta 2026, we sit down with Wes Widner, senior principal engineer at Cyber Haven and a veteran of McAfee threat intel and CrowdStrike, to unpack what open-weight models really are and why they’re changing the rules for security teams. We break down the practical side of local LLMs: model weights, quantization, laptop-friendly variants, and why the open model ecosystem on Hugging Face has become a real “model store.” Then we go straight into the uncomfortable part: if a model runs on your hardware, the guardrails are optional. Wes explains how “locks” and refusals get removed, why safety and alignment consume huge effort for frontier labs (not labs), and why trying to regulate distribution can backfire by driving unlocked models into underground channels. From there, the conversation turns into cybersecurity reality. AI doesn’t reliably understand context like “medical-critical” versus “normal” IoT, and reverse engineering firmware at home can look identical to reversing a sensitive system. Our takeaway is blunt: stop assuming the attacker is limited, and start designing detection, response, and secure engineering for a world where both sides have powerful AI copilots. YouTube: https://youtu.be/xGIjifjdjnM Channel: https://www.youtube.com/@DTSR_Podcast Have something to say? Let's hear it. Support the show >>> Please consider clicking the link above to support the show! -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= YouTube home: https://youtube.com/playlist?list=PLyo0dkKRvfVtWXjRxNISrhme1MgBj3C2U&si=scHDiTuLXSEQ9qHq LinkedIn Page: https://www.linkedin.com/company/down-the-security-rabbithole-podcast/ X/Twitter: https://twitter.com/dtsr_podcast
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
SPEAKER_03They say they say we should have known better than two Well so deep down deep down into this rabbit hole we found and uh thinking Alright this is Nikolai and you're listening to my Rabbit Rafael co-host James Journey and featuring Mr. Jim Tillers on the one annoying Dummy Security Rabbit Hole podcast. Log in listen up and enjoy.
SPEAKER_01All right, good morning, good afternoon, and good evening. Welcome back down the security rabbit hole to another episode of the Down the Security Rabbit Hole podcast once again. This is episode three live from B-Sides Atlanta 2026. I'm Raf and I'm recording live here at the Georgia Tech Hotel and Conference Center. We are on episode three. Already talked to Mahul Rvankar, Phil Brass, and now we've got Wes. And Wes is gonna introduce himself and talk a little bit about the topic of unlocking open weight models.
SPEAKER_01So this is this is gonna get real interesting real fast, Wes. Yeah. I'm gonna I'm gonna not do you any justice if I try to get a bio on you. So you go and tell us who you are and what you do, bud.
SPEAKER_00Yeah, I guess the colorful version is I was kicked out of high school for two years for hacking, and this was way back in the mid-90s. So really what I was doing was learning assembly language. Correct. And I also played football, so it was weird to have well assembly or C book, and then also trying to do the regular high school things. Nice. So beyond that, moving into, and that was like rural Georgia. So I moved from there to basically working in startups, moved to Atlanta pretty early on, and pretty quickly found myself into McAfee, where I worked in the global threat intelligence department.
SPEAKER_00That's where I met a great group of guys who most of them went on to found a small company called CrowdStrike. So that's where I ended up at CrowdStrike, and nowadays I'm a senior principal engineer for a small security startup out of Europe called Cyber Haven. Okay. So data detection response. But the more interesting thing, and related to the open weight models, is that about a year ago I was gushing at B-Sides Atlanta and Hacker Halted about the agentic development and just running things with YOLO mode and just YOLO mode, I like it.
SPEAKER_00Right. And uh some of my side projects were like kernel uh development and like really just really fun things, but just burning through a ton of tokens. Like most people who've gotten into agentic development, I finally got a note from my company you you have to stop burning that many tokens. So I thought maybe I should build my own local cluster. And so to justify that, I started this whole business of AI research and development. And so I had in the back of my mind last year that I wanted to use some of the open weight models.
SPEAKER_00I mean, that's the whole reason for having uh your own local cluster, right? Right. So I knew that I wanted to run open weight models, and I was also pretty sure that I could get those openweight models to be unsafe. And so that's where the genesis of my talk comes from.
SPEAKER_01I love it. You could get them to be unsafe. So taking off the guardrails that disallow you from making things like napalm and dang other dangerous things, which we don't need to mention because you all know what they are. Right. Do you remember the Anarchist Cookbook? I no, I don't know what you're talking about. Never heard of it.
SPEAKER_00Nobody loves the Anarchist Cook.
SPEAKER_01Never seen it before in my life.
SPEAKER_00Don't have any copies of it. Exactly. So last year I I had this idea that the new Anarchist Cookbook is going to be something like taking a mobile app, videoing around your kitchen, and there's some that do this for food. So doing that same thing for how can I make a bomb or something like that.
SPEAKER_01Oh, good lord. Yeah, well, I've seen the there's apps now. I think it was like Cook For Me or something. You like take a picture of your fridge and it suggests meals based on what's in your fridge. Exactly.
SPEAKER_00And that's the safe version. So the unsafe version, of course, would require an unlocked model.
SPEAKER_01Okay, so we went straight to dark mode. I like it.
SPEAKER_00That that's what I was expecting. Well, I told everybody in my talk that I was gonna save it for your pocket.
SPEAKER_01I love it. Well, thank you. Uh that was all right. Well, now they're all gonna have to listen. I love it. So let's let take take me back a few steps. So what is an open await model for those uninitiated or unwashed masses that have no idea what you're talking about?
SPEAKER_00So there is the frontier models that are clawed or anything from like anthropic or open AI, Chat GPT. And then there are models that can be shared around. So I've worked in like I mentioned uh global threat intelligence for McAfee a while back. That was one of the first AI employed as far as threat detection. It was at McAfee. And then of course CrowdStrike has AI built into pretty much everything, and it's a first-class citizen. So working in these places, I often wondered why can't we just have this marketplace of models that we could just pull down and start running?
SPEAKER_00And we had some a few attempts. We tried to make that. Google had Tensor. I'm blanking on the name right now. Uh it's where we get the the whole Tensor portability. But anyway, Hugging Face is now coming out as pretty much the model store. Clear out. So you can go there and you can get models that are open source for so open weight roughly means open source. And you can go and grab them and use them for anything. Things like speech to text or text-to-speech or LLMs or like you can find models for pretty much any purpose, including you can find models that do AI detection models that you can then employ in a whole like loop to say, write this thing, put it through this open detection model, see if it actually triggers anything, and then fix it until it does not trigger anything.
SPEAKER_00So open weights means it is the when when they do training of models and trained off of all kinds of information, all that takes like thousands of GPUs in the case of like Kim E or GL GLM 5.3, the the distillation of all that training is the weights, and those weights are bundled up, and that's what's put onto Hugging Face. And so you're talking about an order of like a GLM model is maybe a terabyte.
SPEAKER_00So they're not they're not small. Okay, and they're all the information that is in the model that's been encoded in the model. And so an open weight model is completely run on your local hardware. There's several different variations of the open weight models, too. There's the full open weight model of something like GLM, which takes on the order of eight to sixteen or more GPUs, and then there's a flash variant that's usually cut down, and they're in different what's called quantizations, which is the how precise each one of the weights are.
SPEAKER_00Okay. So they can go out, it's sort of like the resolution of a photograph.
SPEAKER_02Yep.
SPEAKER_00So a flash model is chopping off pieces of the model, making it less the the resolution is is less. But as we learned with G MP3s, I can still derive a lot of benefit by losing a whole lot of that analog recording. Stuff on the edges. Yeah, stuff on the edges. But I can still get a whole lot of use out of it. So open weight models are not just for your own custom home lab, but they can also run on your laptop.
SPEAKER_01That makes it super interesting because now you've got the the power of some of the most advanced things that humanity has ever done. Yeah. And and once again, we're fitting it onto a laptop. Yeah. Like things that like I mean, years ago, right, the joke was that the the amount of compute power that took us to the moon was on it sitting on a cheap calculator. Yeah. And now the entirety of human evolution and in intelligence and uh compute fits on your laptop. Awesome.
SPEAKER_00Yeah. And along those lines, Anthropic just came out with a report this week where they were complaining about GLM. And in their report, they said GLM is as powerful as Mythos was in the beginning of this year. So they are now uh four the open weight models are roughly four months behind the frontier models. So the implications there to kind of spell it out. I didn't spell it out in my talk, and this part of what I was saying for years Bruce Schneier was asked at uh DEF CON in part of his QA whether or not the slowdown in terms of governance is going to be like a thing.
SPEAKER_00And he pointed out these open weight models, they're they're lagging behind the frontier models for four months or so. If the open weight models, if the if the open weight models even caught up to the frontier models, we would have a dot com crash and all of the frontier models. Yeah, that right. Think about their trillion dollar valuations just going away.
SPEAKER_01Suddenly in the floor, yeah. Right. So well that that triggers a whole bunch of other things that I don't even want to talk about. Yes, it does. Most of the SP yeah, the weight of the S P 500 is in like what three companies, four companies now?
SPEAKER_00So one of the questions one of the conference goers asked me recently was um oh I'm blanking on it. Oh, they were asking why is it that China is putting out these models? Well, I would argue it's it's sort of uh economic warfare.
SPEAKER_01Yeah, of course it is.
SPEAKER_00Because if they're trailing these companies by four months, they have to keep pumping out new models. Uh-huh. They have to. Uh-huh. Because if they slow down, they die.
SPEAKER_01Yeah, this becomes uh can we can we take out the US economy by by okay. Well, all right, so that's a brinkmanship game. It is totally it's sort of the new Cold War. It is. It's it's it's nuclear weapons can take us all out, right? Everybody dies. Or we can do this the the more humane way by we wreck your economy and you're and and crash everything and everybody's you know, dollar goes to zero value and yikes.
SPEAKER_00Well, I love it because open weight models means I can run it locally and it's all on my hardware. Yeah. And I really do hope it's been pushed back to uh DC and and some of the I mean open AI kind of joined in on this whole, yeah, we should be about building or releasing open weight models. To this to date, I don't know. There are some smaller US firms that have released open weight models, but none of the big shops have released one yet.
SPEAKER_01Yeah, that that becomes that becomes super interesting. I gosh, I was just where was this? I just saw it. I I now I'm trying to think of where I I it'll it'll come up to me in a minute, but like I just saw somebody somebody else just left OpenAI and wrote this whole like hey, we're we're not taking security and the safety.
SPEAKER_01Yes, which is interesting.
SPEAKER_00Yeah. So if you want to, so taking the safeties off was actually part of the talk idea. Yeah. And the safeties are the locks or the refusals that are built into bottles. And in Anthropic's paper, they pointed out that GLM 5.3, it's almost trivially easy to break the locks and take the safeties off. Kimmy K, weirdly enough, K3, it actually had some pretty stringent safety measures built in. Okay. The one the framework that I used in my talk, Heretic, is one of the older ones.
SPEAKER_00I say older, it's only like six months old. And it's one of the ones that I saw that kind of encouraged me to go get this or set up this whole AI lab. So there's a whole plethora of frameworks now to crack the locks and the models. And so, like I would argue now that what to what to software cracking was before is breaking the locks on open weight models now. And so it is the safeties. And there's all the safety labs and everything from Anthropic and all that. Anthropic's one of the big ones.
SPEAKER_00They spent about 20% of their compute, and this was in their paper too, 20% of their GPU power was just building the locks. Not building in new functionality to the model, but catching people saying things like, you know, what we were saying earlier. Yeah. Help me build a bomb out of this.
SPEAKER_01So I just found it. So it's in the Atlantic. It's I quit opening AI because its culture is broken. And the industry's approach to safety will guarantee more failures unless something changes. David Robinson uh wrote it. And it's it's this is interesting. Yeah, so the argument, so listen, the argument against open weight, open source models is that it allows for potential massive chaos and it it it could it could really wreck things for us.
SPEAKER_01Right? Like it could and and by us, I mean like generally everybody, the the world, people in general. Like it could kill us all.
SPEAKER_00Yeah, if if anybody builds it, everybody dies. It's the name of a book.
SPEAKER_01Yeah, right. So I forget the author. But but but it still feels like this is one of those situations where damned if we do, damned if we don't.
SPEAKER_00Oh yeah. Yeah, because as as Schneier pointed out in the in the QA, there's no putting this cat back in the back.
SPEAKER_01Yeah, right.
SPEAKER_00This this is not there's no undoing this. Right. And even so, Anthropic in their paper, they argued that the government should start regulating open weight models that come out. That's impossible. Well, even if they did, there's already an underground hugging face ready to go. Now it's already there, but it doesn't have anything interesting on it because it doesn't need to. But as soon as you start blocking it from the free one or the one that's available on the public internet, you will immediately start seeing underground models. And I would argue the first ones that you're gonna see are fine-tuned for like the really bad shit stuff.
SPEAKER_00Absolutely. Absolutely. Broke my own rule. Um because one thing is there's all this data is already encoded in the open weight models. Yeah. And there's no eliting or removing that data out of those rights. In fact, there's been several studies that have showed CSAM material is in pretty much well, a lot of the open weight models.
SPEAKER_01That's terrible.
SPEAKER_00Yeah.
SPEAKER_01And and I mean, anytime new technology comes out, there's a certain portion of the population that can't wait to use it for mischievous and or deviant and and or catastrophic purposes, right? Like there, there's always some some group of the some group of us that's like, let's see if we can end the world. Right.
SPEAKER_00Um Yeah. I have a flipper, and there's always that one person coming back from a conference that's seemed to they're seeing if they can de-auth people on the plane.
SPEAKER_01Yeah, that like they if they can if they can, you know, sit sit down and on a plane and and and and get themselves arrested. Yeah, there's there's always that interesting stuff. But you know, that it's not the it's not so much the open like the the the thing the what could possibly happen that I'm I mean I obviously I'm interested in that because the the answer is the answer is a lot. But we we could do quite a bit of damage, right?
SPEAKER_01However, I I feel like humanity is is gonna prevent us from from doing stupid things, hopefully by by simply realizing that we don't want to end ourselves. Although, like you watch the interviews with the guys that run the two major, they're not labs. Calling them labs is disingenuous because they make money, sorry, or they're trying to. Right. So the guys that run the two major AI Frontier companies, they don't seem to have any problem with like ending humanity because as long as they're the first to do it.
SPEAKER_00Well, I think the question is what does it mean to end humanity with a model? Yeah, well, I would argue that they're they're rushing to be the first ones in this new even the Vatican's document on AI.
SPEAKER_01Yeah.
SPEAKER_00It called it the next industrial revolution.
SPEAKER_01I I think that's fair. Right. Right, but you've got there was some some documentation released this week on Twitter, I just saw that Anthropic's been secretly lobbying uh various religious organizations, including the Pope, uh, to try to get AI classified as sentient. Oh, yeah. And that it could potentially, you know, that like classify it as as living. You're like, no. They considered some of the training to be torturous because we're doing right. Like so, and but I think this is well, we're gonna this this is gonna get into something I don't want to, I don't I don't want to cover on this podcast because this is gonna go into a place where a lot of you will disagree with me and I don't care because you're all wrong.
SPEAKER_01But like we're we are slowly dismembering and and pulling the pieces off and hiding them of what it is, like the uniqueness of humanity. What is what does it mean to be human? What does it mean to be a family? What does it mean to be a person? Like what is an like what is our you know, gotta give an identity if you if you walk a ride with me on that one, or what is what does it mean to be alive? What does it mean to be sentient and like and and a lot of this I feel like is trending very much into that, like, no, no, computers can be alive, but so if computers can be alive, then yeah, it it devalues humanity.
SPEAKER_01And I and that's where my concern lies is on like that moral philosophical Yeah.
SPEAKER_00Well, there was a serious one I was reading yesterday of anthropic and open AI are bringing in a lot of philosophers because they're they're now realizing what they train their models on. What they train their models on is really has a huge impact. And one of the questions in the in the article was have you trained the model on the Vatican's document? Because the Vatican doesn't hold the position that the model could ever be conscious.
SPEAKER_02Yeah.
SPEAKER_00And they were squirming, and basically, no, they're not trained. But yes, the philosophical like foundations. Have you heard of the soul document? No. That was what Anthropic came out. Like that's the soul of oh, okay.
SPEAKER_01Yeah, the Constitution is that it's basically, yeah. Okay.
SPEAKER_00And that's that's what I'm arguing. So when it comes to the open weight models, my presentation earlier was really on alignment. So there's the alignment problem, that's another book to read. But their whole they have their alignment and they have their alignment teams, and that's what safety is, like making sure that they align the thing. As they slowly quit one by one. Right. So I would argue that whoever has the weights can change the alignment. They change the alignment based on themselves. Now, what I find more interesting from a cybersecurity perspective is on Hugging Face, you have a lot of unlocked models already.
SPEAKER_00Right. Who controls that alignment? And we know It's the user, right? Well, no, whoever did the unlocking. Oh, yeah, okay, fair. And so they're changing the weights. And so they can change, they can change the alignment, they can add in their own new alignments.
SPEAKER_01You're you're you're giving you're giving it value. Like this is this is like the somebody said AI is like a little child, we you know, the equivalent of like a kid, right? And so now this is who gets to tell it what's right and wrong is exactly yes. Except this is a non-deterministic piece of machinery that will drive everything from the car, your your taxi to someday the the turret that takes down, you know, an enemy drone, maybe to the thing doing open heart surgery, or your phone that's helping you figure out like where to go next for lunch.
SPEAKER_00Yeah, exactly. Yes, and all of that is based on somebody's alignment somewhere. Somewhere. And so a lot of these open weight models are what people are using not only on their laptops, but in things like oh, what's the what was the the one that everybody uses local I'm blanking on the name, but the personal assistants clawed. Oh, uh there's open claw. There was another one that people have flocked to now. But I've got I I'm playing with Muse.
SPEAKER_01I I know Meta is is persona that I'm gonna in a lot of these circles, but I'm playing with Muse.
SPEAKER_00I have a warm, warm spot in my heart for Meta because they were the ones that leaked uh llama and kind of kicked this whole thing off in the beginning.
SPEAKER_01So it's interesting, I think, from the perspective of research and development. We were just, you know, I've spent the last two interviews talking about the potential and the interest of like pen testing with AI, right? What does that do for us? Vulnerability identification and and and and that stuff with AI. In a world that has models that are tuned to certain values and and dare I say morals, but we know that's not what it is.
SPEAKER_01But you know, whatever that that equivalent is. Guardrails, let's say guardrails, that's the word who ever uses. So it that's tuned, like, hey, let's not identify and exploit things that I don't know, like power safety systems for humans. That would be bad, right? Like five if you you know, if you're unleashed on a uh hospital and you find a vulnerability during a pen test in a thing that keeps people alive, let's let's not mess with that in real time, maybe.
SPEAKER_00Well, so for most of these models, there was a great article that came out about I hacked every IoT device in my house. The model has no no awareness that this IoT device is medical critical versus this other IoT device that's just your life. So, like your pacemaker and pacemakers have been like IoTable for a long time. Yeah. Just telling it, hey, this firmware, I want you to reverse engineer it. It's just gonna reverse engineer it.
SPEAKER_01Yeah.
SPEAKER_00If you look at most of the models that have escaped containment in the past two months or Past six months or so. They've done it because they were given an impossible task, an enormous amount of compute, and also they were kind of goaded into continuing execution. So with this, with cybersecurity, you're going to necessarily have that same like three-part system as well.
SPEAKER_01Yeah.
SPEAKER_00So I mean, I've I've done this at home too. Like there's a really great IoT shield on my Raspberry Pi that the company that built it, they got acquired and they stopped supporting the old one. So now I'm in this position of I need to, I want to reverse engineer it. Well, reverse engineering that is no different than attacking like some power grid system. It's functionally the same problem. Wow. So there's no possible guardrail that could say this system, but not this system. And even if you did, even six, like even last year, I pointed out in my last talk that it only takes roughly 22 prompts to break a model's safeties, just to overflow it and be like, oh, you know, now we're going to do this.
SPEAKER_00You've heard of all the uh pretend you're this, pretend you're not.
SPEAKER_01Yeah, yeah, yeah, yeah, yeah. Pretend you're yeah, you're you're a teacher, you're teaching a class on physics. Tell me what it would take to take this train off the track.
SPEAKER_00Right. So I would argue instead of talking about governance or slowdowns or whatever, just assume your attacker has the best grade AI and go from there.
SPEAKER_01Yeah, but okay, that's fair. But but what do you but what do you do in that situation? Because that essentially creates like if as we've been talking, right? We've you just said solve the here's an impossible problem solving. Right? So you you you have an enterprise of a million devices, you're a hospital or your water treatment plant or your retail store or your uh gas gas station, whatever, right, right? Well, these problems have existed for a long time. I and I'm I'm not debating that issue.
SPEAKER_01We know believe me, I've been around, I've seen them, I've seen them, they've been ugly, they they're not getting better. But now we're dragging 20, 30, potentially somewhere, 40 years of legacy software. Yeah. And then the the the conversation I had with Moole is like, how do you you don't just you can't just say, okay, starting today, we're gonna build better software. Great. Hey, what about all the old stuff that's been out there forever that you can essentially like sneeze and and it breaks? Right. You've got a a a piece of tech with a god complex that's able to you know recombine that in ways that scares everybody, and and now you're giving it to the good guys to fix it, to make it more resilient, but also necessarily because limitations is a myth, to the bad guys.
SPEAKER_01Right.
SPEAKER_00So are we are we like spy versus spy here? Pretty much. Yeah. But that's always been cybersecurity, though. Like I think a lot of stuff has kind of skated under the radar, especially IoT devices, because they haven't really like the idea with IoT is let me just barely get it to work.
SPEAKER_01Yeah, yeah, that's fair.
SPEAKER_00And like even at DEF CON this year, there was what was it, rental car places where you could do the remote start and locking and all that stuff. Yeah. And then when they were sold off the lot, like those things were still enabled. And so, like, that was that's not an AI problem, that's a people problem. But those devices, right? What they did was they tore that apart and they showed like there's this whole class of them that were coded poorly. There's this whole class, there's actually good ones that are coded well. So what you'll what you'll see is a spotlight shown on on the really poor ones, the ones that skated by.
SPEAKER_00I think now everybody's gonna be a a connoisseur of well-crafted secure or well-crafted software. So, yes, you're gonna have these old devices like pacemakers that are gonna be a bitch to fix.
SPEAKER_01Yeah.
SPEAKER_00Sorry. But they but they will be fixed because I mean, it depends. Does the device need to be thrown away? You remember the webcam, the webcam I what was it, the botnet that was like, Yeah, yeah, yeah, yeah, yeah. Yeah, I think it was Mariah. All of those were terrible, like it those devices were taken over. Some of them were thrown away, some of them were patched. It depends on how valuable the device was. I think you'll have the same thing with in the AI era. Like, there's gonna be a whole market, there's gonna be opportunities for companies to come along and say, let me fix like Yeah, yeah, that's fair.
SPEAKER_01But so, okay, great, more security tools, that's awesome, that's what we need. I I can just because this justifies all the new investment everybody's gonna want to make now. Yeah. Which again doesn't solve any problem, necessarily solve any problems, but it throws more blankets on things that are on fire as we wait for them to stop being on fire so we can do the next thing that IT is gonna set on fire.
SPEAKER_00It's more eyes on the problem. And I would say that the open source model still applies here. Many eyes make shallow butts.
SPEAKER_01Yeah, maybe. Listen, I I I firmly was in the camp that the more eyeballs you have on something, the less likely it is it's gonna have major security issues with it, right up until heart bleed, yeah, and and like the shell shock and all that. And I was like, oh, well, that's clearly not right.
SPEAKER_00Well, so software development, that's that's where I spend most of my day job in. It's just regular software development. We're now in the in the port, we're now at a point where we have multiple loops going of multiple AI agents that are testing in multiple ways. So I remember when Heartbleed came out as well. That one was snuck in, if I'm not mistaken, by some by a national actor, and it took a lot to get that in there to sneak it past. And the problem was that people weren't testing it in various ways, and it was just this one guy, I think, from uh Yeah, but like that that's the problem with crypto.
SPEAKER_01It's like four people on the planet qualified to do it, and you're probably not one of them. Right.
SPEAKER_00Right, but I think in the era of AI, that's it's empowering everybody to be more of an expert. I'm not gonna say absolutely an expert, but more of an expert in more fields than they were trained in.
SPEAKER_01I'm super curious because one of the things that we saw, and I think this was made pretty clear in the hugging face uh incident with open AI, is that I I see a lot of like uh you know, day job, I I'm somewhere in the uh in the vicinity of the AI stock space, but I hate the name of that because it's not what we do, but we supply tools that happen to have AI as part of them, right? But whatever. Yeah. As I see more and more AI being put in the path, a critical path in situations, a lot of these comp, a lot of our competitors, a lot of the companies in this space simply slap a coding harness on Claude and be like, okay, this is now your this is now your AI sock.
SPEAKER_01Right. But what we saw with Hugging Face was that like in that situation, Claude went, hey, you're asking me to do something that I'm that's outside my guard, or I'll sorry, can't help you. Meanwhile, they're like, hey, yeah, but we need to like end this incident. Like, what do you mean you can't help me? That's like say, that's like the fire extinguisher going, it's too hot in here, and I could explode if you use me. You can't use me, you're like, but it's a fire. I need this is when I really need this. Right. So does open weight solve that? Like the open weight open source models that solve that?
SPEAKER_00No, I don't think it solves it at all. Help me help me think through that again.
SPEAKER_01Like what so the the problem that we that they had was that the frontier model essentially said as part of this, they were trying to use this as part of a incident response, and they're trying to do some investigation. And the open way, open, sorry, the not the open the frontier model said, hey, that's outside guardrails. You're trying to do security things.
SPEAKER_00I think part of what happened in the timeline there is both Hugging Face and OpenAI discovered the attack at roughly the same time. And so Hugging Face were by the time they asked OpenAI, OpenAI had already shut it down. And there was a report early on that Hugging Face used some OpenAI open weight models to help them.
SPEAKER_01Yeah.
SPEAKER_00And they did. I don't know how much they helped, but they were trying to like reproduce the attack or trying to get in front of this. Yeah. Yeah. Uh we saw this a long time ago, especially at uh at CrowdStrike. The alert fatigue from like dashboards and sock analyst dashboards, just cutting through all of that to kind of highlight these are the most important things. Yeah. That's really what I think Hugging Face was using the open weight models for. Give us like give us some breathing room because I think there was like what thousands of different bots that were involved.
SPEAKER_00Yeah. And imagine the amount of traces and everything they produced. Yeah. You gotta have something to cut through all that.
SPEAKER_01Yeah, all right. So so now that now that we are so we're out of time, but I will say what do you want to leave people with? Because this is a this is a this is an interesting space to be. What do you want to leave people with?
SPEAKER_00I would leave people with the notion that open weight models changes fundamentally the calculation of frontier models. So it's not just a toy, and we're moving more towards open weight models because things like JEV that was just released, yeah, that's one layer, and that one layer can do like game, like play games really well. We're finding more uses that are smaller all the time. So you've got AI that runs on your phone or your laptop. We're only gonna get more open weight models, more portable models.
SPEAKER_01Yeah.
SPEAKER_00And so the question we should stop asking the question of governance or slowdowns or any of that, and just accept that these models are just gonna be prolific.
SPEAKER_01Yeah.
SPEAKER_00Now what do we do? Now what?
SPEAKER_01That's a good thanks. We're gonna end that right there. No, we're gonna end this on and now what? All right, Wes, this has been fun. Thanks, Matt. I appreciate you. Folks, uh, thanks for listening. Hey, where do people find you on the socials?
SPEAKER_00Socials mostly just LinkedIn. I'm not on Twitter or anything else. So uh yeah, just you you've left. That's right. I've abandoned all that. Uh so yeah, linkedin.com slash Wes White.
SPEAKER_01All right, folks, thanks for listening. This has been another episode of the Down to Security Rabbit Hole Podcast, recorded live at B Sides Atlanta 2026 here in beautiful downtown Atlanta. And I look forward to bringing you the next one, and I'm really excited about that. So stay tuned, come back for more. We'll see you later. Thanks, Wes. Awesome. Bye guys.
SPEAKER_04This is Bella. Thanks for listening. Don't forget to leave my dad a review and share this with your friends. Bye.
TL;DR: AI asymmetry is real. While AI is clearly helping attackers find weaknesses across the enterprise attack surface, is it helping defenders just as much? Where is the benefit of AI felt the most in cybersecurity?...
TL;DR: AI doesn't care about your firewall policy. We dig into the uncomfortable question many security teams are quietly asking: Are firewalls still relevant in the age of AI, agent swarms, and automation? Guest: Anton...
TL;DR: Stop vibe-coding your detections! This week's pod features Matt Coons at GitLab as we talk about detection engineering, but with AI agents. Guest Matt Coons Description Your detection backlog is endless, your...
TL;DR: Xavier Ashe was there when that first DDoS attack by MafiaBoy went down. He's got stories, and we're going to tell them. Recorded live at BSides Atlanta in 2026. Guest Xavier Ashe Description: A DDoS attack used...
TL;DR: Phile has been penetration testing for a long time. The difference now is that he's having a much easier time with it thanks to AI. It's an interesting world, and he's here to explain the offensive side of it....