Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

The Low Down

The Low Down artwork

Live from DEF CON: AI Harnesses, 1000+ Linux LPEs, and the Shia Haloud Supply Chain Worm Returns

Aug 12, 202641 minAudioSeason 1, Episode 7
ShareXLinkedInFacebookWhatsAppEmail

Welcome to The Low Down, the best show on the internet for hackers The Low Down is presented by Maze. LinkedIn: https://www.linkedin.com/company/mazehq/ X: https://twitter.com/Maze_Security Follow Us! https://www.instagram.com/lowdown.pod This week we're broadcasting live from DEF CON 34 and Black Hat, bringing you the most important conversations happening at Hacker Summer Camp about AI security research, vulnerability orchestration, and the evolving threat landscape. Today we're talking about: Live from DEF CON: The Community Experience Recording from the Biohacking Village at DEF CON 34, we discuss what makes this conference special. From the Maritime Hacking Village to the DDoS Village, we explore how the grassroots community atmosphere differs from Black Hat's corporate environment. Meeting listeners, getting recognized by village organizers, and why the code word is potato salad. Black Hat Keynotes: The Future of AI Vulnerability Research Dave Weston from Microsoft and Professor Jan from ASU delivered game changing talks on what happens when AI makes vulnerabilities abundant rather than scarce. Breaking down how ASU accidentally proved that orchestrated Codex agents outperform single Mythos instances, and what this means for the economics of vulnerability markets and formal verification. The Harness Revolution: Why Orchestration Still Matters Deep dive into why we're in a sine wave pattern between model capabilities and harness quality. The real juice isn't telling AI to find bugs with no mistakes. It's building sophisticated vulnerability research machines that leverage AI as one component. Why context management and scaffolding remain the competitive advantage even as models improve. 1Password Research: The Auto Remediation Reality Check Keith's team at 1Password drops research showing only 26% of AI generated vulnerability patches were actually usable. Breaking down how even the best models partnering with top tier firms like Trail of Bits still introduce new bugs while fixing old ones. Why humans in the loop remain essential and you can't change the shape of the pipeline that produces bugs just by patching faster. Meeting OpenAI Leadership: The Cyber Model Question Conversation with OpenAI co founder Greg about what would happen if they released their most capable cyber model with no guardrails. Discussing the philosophical questions around AI security research capabilities, export controls, and whether we're measuring danger correctly. Plus OpenAI shows up unscheduled at Black Hat to explain the Hugging Face incident from their perspective. Black Hat Vendor Floor: The AI Crab Evolution Examining how the security vendor landscape has evolved into 75 plus AI SOC companies all converging on similar solutions. Why evolution keeps producing crabs in nature and in cybersecurity products. The challenge of differentiation when investor pressure pushes everyone toward the same features. Casey Ellis and Disclose.io: Fixing Vulnerability Reporting BugCrowd founder Casey Ellis launches Disclose.io to solve the full disclosure problem plaguing smaller vendors. How an AI agent runs nightly to self heal gaps in the vulnerability reporting database. Why researchers claiming they couldn't find who to report to is often BS, and how this project aims to remove that excuse. Jason Haddix Revives BEEF Framework Arcanum open sources a modernized version of the Browser Exploitation Framework that was abandoned years ago. Why browser based exploitation tools still matter for red team engagements and what the new BEEF brings to pen testing workflows. Shia Haloud Supply Chain Worm Returns The self replicating NPM supply chain worm strikes again, compromising packages with 2 billion downloads per month including Key V, flat cache, and file entry cache. Breaking down how the worm steals NPM tokens, GitHub credentials, AWS keys, HashiCorp Vault secrets, Kubernetes configs, and AI service credentials. Why the second order effects of all these stolen secrets keep security teams up at night. Iran's Water Supply Attacks and the PLC Problem Following up on Iranian threat actors targeting Minnesota water infrastructure through internet exposed PLCs vulnerable to CVEs from 2013. General Nakasone discusses the impossible scaling problem of securing 50,000 water municipalities across the US. Why there's no bat phone to call when critical infrastructure is vulnerable, the CISA defunding irony, and the knowledge management failures that plague operational technology security.

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.

Good morning, hackers. Welcome to the Lowdown presented by Maze, our code security partners for this podcast. Guys, I don't know if you can tell, but we are not in my basement. Matt is not in his studio, and we're not at PlanetSkale headquarters. Guys, we are at DEFCON. We are at Hacker Summer Camp having a great old time. Matt, how's your Vegas experience going for you so far? I got to say, your boy is not 19 anymore. Yeah, I did the thing I always do. I get to Thursday and I changed my flight to a day earlier.

So I, as you can hear, my voice is maybe at about a 40 out of a hundred right now. So love you for sticking through this. I also just want to say thanks to the biohacking village for hosting us here, the WeHack Health crew, because we were having a hard time finding a good spot where we could film this because DEF CON makes it, you know, so that not everyone's just. in backgrounds of photos and all this kind of stuff, which is great. It makes sense. There's people here that don't want to be seen and on the internet and all this kind of stuff.

We actually got to go to dinner last night with a bunch of creators. We took over a whole Korean barbecue restaurant. Shout out Sean from Miscreants for setting that up. But, like, case in point, Jack Recider from Darknet Diaries was there. Yeah. Not in photos. We took a big group photo, and he, like, scurries away. He's there. We have proof, sort of. Yeah, sure. Like, he gave me a little bracelet. Yeah, man. But, yeah, the homies are out, you know? So it's really cool to be here another year. I've been coming every year since, like, 2009. What was your first? Oh, first was DEFCON 21. What is that? 2020 or 2013.

Yeah. Yeah, yeah. That's only a few years after me. That was actually the first time that I did the Black Hat DEFCON double headers. If you're not aware, Black Hat is kind of like the more corporate, a little more sales heavy, rightfully so, different kind of organizations. Conference before DEFCON, and then DEFCON happens after. And so when I was 19 at the time in 2013, you know, I was a little more spry and I could do that. No issue. I did the back to back this year for the first time in 13 years. And I am I'm fried. Yeah, so my first year, I think I did Sunday to Sunday, something like that.

But then pretty quickly, I actually spoke in 2011. So like 2009, my first ever thing. I got my talk accepted 2011. Black Hat, the second ever B-Sides LV, maybe third. And then DEFCON that Saturday. I had to give the same talk three times back then. I will say it is much easier to get through a full week in Vegas not drinking booze. here yeah like it's seven days like vegas while drinking is a three-day city yes i totally agree um it's nice i highly recommend if you're in the industry and you haven't gone out in here yet at least do one defcon man like everyone here is so nice and it's gotten so big where there are villages for literally every sector of cyber security if you're an embedded person if you're a red teamer if you're a physical pen tester you like to break down doors there's literally a spot for you not an ad just literally it's that it's that great so come hang out

We're literally in, we said we're in the biohacking village. We're in the village hall. We've done a couple of laps already this morning. It's my favorite thing to do at DEFCON because these village organizers are so happy to be here and so passionate that you're at their village checking their hard workout. Right? So the maritime hacking village is like flagging people in and they're like, come on, hack our boat. Check out this drone ship that takes down drug mules. Right. You know, off of Florida. And, like, they're so excited you're there. They wanted to show you what was going on.

We walked into the DDoS village. They're really excited to see low level because one of the routers that they have there, the only reason they have it is because they watched your video. Right. Yeah, yeah, yeah. Like, it's so, so, so cool. Well, A, because we're who we are that, like, you know, some people kind of have the parasocial, like, oh, cool. I'm excited. I've said hi to so many awesome people. It's not even halfway over. We're only, like, this is day one of DEFCON. So, yeah. Yeah, I'm going to give low levels some crap because they only recognize him, which means he's not promoting the podcast on his videos enough yet.

Come on, man. Come on. No, I'm just messing with you. Yeah, we have very different audiences. Black Hat, I was getting pulled over left and right. People shoving stuff in my face. But that like... It's just so cool to get that. Because we're just screaming into the void. Even here, we're recording live. No one's watching us. We're chilling. I mean, biohacking homie who's running the audio is watching us. But we scream into the void a lot in our house. It's really cool to come here and talk to you guys who actually watch and listen to this. Get feedback, good and bad. And mostly good, honestly.

People are just stoked that we're around. Yeah. Not just that, not just like stroke our ego stuff, but like they literally only had that router on that table because they watch your video. Yeah. Yeah. I was telling people that come up to me, it's like content is such kind of like a one way thing. Like you're like, oh no, surely we comment on your videos. But it's like for us, we put the video into the machine and like that's the end of it, right? So having people come up physically and be like, hey, I watched your video and like I got this job because I brought it up in an interview or like, oh, I found this bug is the way you mentioned this thing. Like it just, it's really, really nice to hear that.

DEFCON is a very fun community event for us in that regard. And then this is the only place that the kind of stuff that happens. Like I was at a party last night for like an old like threat intel sharing malware email list that I've been in and out of and stuff. And they still run a party. And I had people coming up to me saying, hey, I run the SOC for this whole state. And we share your videos around the SOC. That's how we find out about some stuff. My boss wants to meet you. Can we hang out tomorrow?

I had a CIA agent come up to me and be like, hey, man, love your stuff. And I'm like, what? That's crazy. That's crazy. okay cool CIA is on Instagram I guess you know whatever it is so anyway just shout thank you guys for doing that thank you for coming up and like we even said it just now walking back from lunch that there's probably only a small percentage that actually like take the time to even say that to us that you know might recognize us walking around it's just we're grateful for it so thank you appreciate that The other bit, right? So DEFCON and the villages are really where it's at, and it's super cool.

But you also got to meet a general this morning because we have press badges. Yeah, so interesting. I have a press badge, and I'm like, oh, what's on the press agenda today? And it was like, oh, General Nakasone, former director of the NSA. I'm going to go to that press meeting and see what he has to say. It was a conversation about, like, what does AI mean for national security? We talked about, like, the whole open way conversations. It was interesting to hear, like, what he has to say with his perspective on it, right? So super interesting there, for sure. Again, just things that only happen at DEF CON. Yeah, only at DEF CON, baby.

This kind of stuff. For me, it was earlier in the week, Black Hat. So DEF CON's really this, the community. It's really, really community-driven. You can tell. Even, like, 34 years in, it's very grassroots. All the villages are run by the village organizers. It's not like some corporate entity that's putting on these villages, right? Black Hat is very corporate entity, and it feels like it. This is like, whatever. I'm not even saying that's an awful thing. It just is what it is. Different, isn't it? But the events at Black Hat are all by our corporate partners, friends of ours that run businesses or sponsors of ours, or people that we know that work in the industry that invite us to these events.

And so, yeah, I'd say earlier in the week, it was very corporate event heavy, and now we're like, in the community, all this kind of stuff. I will say it was pretty cool earlier in the week. A mutual friend of ours, Clint Gibbler, who runs TLDRSec Newsletter, now at OpenAI, got to invite us to the OpenAI party, and I got to meet the co-founder of OpenAI and talk to him about this kind of stuff. He asked me directly, he was like, what would happen if we released... The cyber model, no guardrails, most capable cyber thing we could produce. If we just released it tomorrow, what would happen?

And we had like a long conversation about this. The co-founder of OpenAI. Again, this is only Black Hat and DefCon. Only Black Hat and DefCon, baby. And a little teaser, we are trying to get him to come talk to us. on camera about that. Like, I don't know if it'll happen. No promises. Future episode. Hit that subscribe button. Yeah. Yeah. Yeah. But like he did, he was like, I'm down. I want to keep talking about this. Cyber is super important. Right. You know? So Greg, thank you. If you're, if you're, you're not listening to this. Greg, thank you. I know you're the biggest fan. Thank you for taking the time to talk to me over there.

But, yeah, that kind of stuff is just great. Did you have any highlights? You said you haven't been to Black Hat in quite a while. A long time. Did anything stand out to you? Obviously, it's way different. A lot of AI companies. A lot of companies getting real deep into AI. Which is like completely expected, right? Because that is where the industry is going. And I totally get it. It just makes me begin to ask, like, how many products are there left to build? You know what I mean? Like, what else are we going to make here, guys? The joke that I've been making, there's this word. I shouldn't. I've said this line so many times this week.

I should look up the word. There's a word in science about how evolution keeps producing crabs. Like in the tree of life, crab-like things have evolved independently of one another so many times because it's like a very efficient body or whatever it is, right? I'm not whatever. Sure. Okay. Cool. I said, that's happening now. Like two years ago, AI was in the startup village, right? And like, maybe it was on a few like posters or whatever, but it wasn't like the headline in any of the big booths. Last year, it was like, okay, it's basically all the startups and some of the big booths are definitely, it's percolating, it's percolating, it's percolating.

Some of the, some of the big AI companies caught some big VC checks and got booths and stuff like this. This year, I think it would be easier to count the booths that didn't have AI as a headline than it would possibly do it, whatever. And a couple years ago, it was like, this is how we're trying to be different. Like, oh, no, we're not just an AI sock because that already exists. We're doing this niche thing over here. And I'd say the vast majority now is like evolution has produced a lot of crabs over here. Like they've all kind of migrated towards the same company, probably due to like investor pressure is my gut.

Right. Yeah. If they're all VC funded, the VCs are like looking at some of the big A.I. socks going, well, why aren't you doing that? You know, kind of thing. And no, no, no, no. We were trying to do this niche. Well, that niche isn't big enough. Like you got to do this. Right. So I could see that happening. Sprinkle some A.I. in there, please. Yeah, yeah. So a lot, I think I heard 75, 77 AI SOC companies. That's so crazy. Yeah, yeah, yeah. So obviously, I'm not saying it's universal. Some are standouts. Some do have differentiators. Some of the earlier players to the game are doing really cool stuff. It's just, it's hard being in the industry, looking at that and going, how do I make sense of this?

Like, if I'm putting one of my old job hats on, how do I pick? Right. It's like, oh, my God. So to kind of transition to the next part of the podcast, one of the highlights that I actually did have from Black Hat on the first of all, the two keynotes at Black Hat were very, very good. The first one was Dave Weston. He goes by Dwizzle. He is a AI VP at Microsoft. And he kind of gave this presentation on like what happens to vulnerabilities, what happens to. exploitation when AI makes vulnerabilities no longer a scarcity, right?

Like what does that do to the market? It's a whole conversation about where the actual businesses are that matter in like the five to 10 year span. And what's kind of cool about the conversation is like what it does lead to is like a market that truly incentivizes the money being in formal verification and memory safety software, as opposed to right now where it's just like reactionary defense type stuff. Right. And then the second keynote, which really interesting, Jan, I don't know his last name, it's like a Russian name, it's very hard to pronounce. Assistant professor at ASU, the author of Anger.

So Anger is a symbolic execution engine for doing vulnerable research in reverse engineering. He gave a really good talk on what is left for the human to do when you're doing vulnerable research now that AI is a thing. And he gave this really cool graph that basically showed what they did at ASU on accident because they actually had unlimited codex spend that ASC was paying for that ASC was not aware of that they were paying for. So like a million or two dollars later, they fixed it. But they actually proved that a single Mythos instance based on the numbers that Mythos reported they found bugs in the Linux kernel is less performant

than a bunch of separate codex agents. So the actual hard problem is still left to the human to do the orchestration and writing the harnesses and to use the Cloudflare harness that we talk about all the time as the magnum opus for the security harness that makes AI's actually good. So there still are products to build, research to do, and problems to solve for humans in the world of orchestration for VR, which is really interesting. In doing that, by the way, they said they found over a thousand non-root-to-root LPEs in the Linux kernel using a tool they came up with.

They're currently in Disclosure trying to figure out, like, what do we do? Because there aren't enough people on the Linux security team. to patch by hand 1,000 LPEs. Do we now Vibecode patches? What does that do for the actual state of the Linux kernel software quality, right? But it was a very interesting opener to an otherwise— not that Black Hat's a non-technical conference, but it's a little more marketing and salesy— to have him be there and kind of anchor it in the reality of, like, this is where the vulns occur, this is where the breaches start, was really, really nice, for more I said. I've got two follow-ups on that exact story.

So I feel like we're actually in a sine wave pattern of this. This was my mantra back two years ago. AI is starting to become ubiquitous. Coding agents are starting to come out. I was saying, hey, our next couple of years, I was calling it scaffolding at the time, because harness wasn't really a phrase that we were using yet. I was like, it's really going to be in the scaffolding. I was using a tool called Klein for a while. a Claude code, like precursor type agent automation engine that was using the models.

And so I was saying, oh, it's all going to be scaffolding and context management was like my sound bite. You can go back and like I said it a million times a couple years ago, right? And then Opus 4.6 dropped. And I was like, oh, the models still have a lot of room to get better. I didn't know. No one knew that. No one realized that the models could still get that much better, right? An Opus 4.6 out of the box, no harnesses, no scaffolding, no context management could do things that the best of the best of the best tools, like award-winning tools.

like couldn't do a few months before that, right? A few weeks before that. And so this is the sine wave. It's like, oh, harnesses, harnesses, harnesses. Oh crap, the models are actually way, way, way better. And now I feel like we're in another part of the sine wave. The harnesses have been the topic of the week for me again, right? Because we're all now realizing the new reality of model capabilities, open weight model capabilities. And again, The real, real juice from the squeeze is, oh, how good is your harness? Right. What can you do?

What can you point it at specifically, like, when it comes to vulnerability research? Or I'm hearing a lot about auto-remediation, right? Like, oh, can I fix things? Not just find them, but fix them at the same time, stuff like this. The second... like story that I heard this week, some research that dropped out of 1Password. My friend Keith started their security research department there just recently, and they just dropped their first research paper. They went and studied a lot of these AI companies that are like partnering and doing a lot of these find and fix vulnerabilities in open source software type stuff, like Patch the Planet, that we're hopefully going to talk to Dan Guido soon about.

Well, 1Password did some research on a bunch of the public PRs for these things. And yeah, they fixed the vulns, but in significant percentage of the time, they actually introduced new problems or broke new code. So it says only 26% of the patches generated were actually usable. And again, the title of the paper is very much targeted towards humans in the loop are still very much necessary in this. This was good news for anyone in this industry that's kind of nervous about all of this, is that even the best of the best, I think Trail of Bits is doing fantastic work, partnering with OpenAI's best models.

These are very, very capable things. There's still cracks, it seems. I'm excited to ask Dan about this. I haven't seen their reaction to the 1Password research. We're seeing that even in the effort with Ubuntu to rewrite all the core utils in Rust. Even if you write... code in a memory-safe language, A, you're still vulnerable to traditional logic errors, cryptographic misimplementation, DOSes through a vector misindexing. But on top of that, I think it was Casey Moratori that said this on the Stand Up, the Primogens podcast. When you have tools that have certain failures in their behavior, they don't do exactly what they're supposed to do,

that is the feature. Like, that is what people get used to. And if you have infrastructure built around those edge cases and then you completely change them, that in and of itself is also a vulnerability. So, yeah, the code is safer, but it's also different. You know what I mean? So it's rewriting or even fixing bugs does not necessarily make everything better net zero. Yeah, this brought me right back to the Cloudflare blog post that I keep bringing up from the CISO during the Mythos Glasswing stuff. And he said this line, he said, the speed at which you fix a flaw does not change the shape of the pipeline that produced the bug.

Right. And this is just yet another piece of evidence to that. Right. We're throwing all of this horsepower at finding and fixing vulnerabilities because we now have like a super coder machine that we didn't have before. a couple months ago, right? So it makes sense, like find and fix all the vulnerabilities. And it turns out we're still not changing the shape of the pipeline that's creating the code by doing this, right? And so again, what is the answer? When I talk to Keith from Unpass about this, guess what? It's harnesses, it's humans, it's all this kind of stuff that I think we're in this sine wave right now.

Super interesting. I had a few other highlights from Black Hat that I was talking to. Some homies, Casey Ellis. Do you know Casey? Founder of Bug Crowd, if you guys don't know Casey. Very passionate about the vulnerability reporting space. Obviously, founder of Bug Crowd, right? Literally tried to solve bug bounty reporting and triage as a career and as a company successfully. Tried, yeah. Yeah. I think Bunker is doing just fine, right? Right, right. Well, Casey just released a new project that he was getting to show me on the side, Disclose.io.

Did you see this? No, no, no. So he did this in reaction to a topic that we've talked about on the show a few times, the security researcher community kind of being a little punks again, right? Right, right. Which I say punk with love, you know? Oh, I did see this. Yep. So he did it in the reaction of like, hey, the security community dropping full disclosure bugs is kind of in vogue right now. Right. We're seeing it a lot. People are frustrated with bug reporting going on at Microsoft or Apple or Google or any of the big shops.

And so they're going full disclosure. Well, that's had ripple effects. And we've actually we haven't really talked about every one of these, but a lot of smaller vendors are getting hit with full disclosure stuff. Just drops on Twitter and GitHub and things like that, because. People are just like, I don't know, getting street cred or whatever they're doing. But they say publicly it was because they couldn't find who to like. Oh, the reporting mechanism of this company was too convoluted. I couldn't figure it out. And Casey's like, I call BS. Yeah, sure. You didn't try. And so Disclose.io is his attempt to catalog.

If you type in a company or a website or something, it will give you how to report a security issue to them. Period. And so the other cool thing is he's a bit AI-pilled. And he's got it, like he showed me a little bit behind the curtain so that every time you type something in and it fails, like if he doesn't have it in his database, that goes in the logs. And every night he's got an AI agent that looks at all the failed queries and then figures out why they failed and tries to self-heal those things.

They're not really bugs. They're just like holes in the data. And so every night he's got an agent. That's a great use for AI, in my opinion. That's what you should be using it for. It's very specific tasks. Right. That's awesome. So kudos to Casey. Super cool. Check out the project. It's open. So I'm sure he would be open for help. It's open source and everything. The other one was Jason Haddix. You know Jason. Yeah. Jadix, a homie. Always great to see him in Vegas. He's the busiest man of the week ever. I think he's speaking at four villages over the next two days. Definitely, he's already given talks earlier in the week.

Jason and his wife gave me this cool bracelet. So thank you guys. It says, hack me, daddy. Let's go. So Jason actually just released via his company Arcanum, which is a red team shop, security consulting shop. They do a lot of cool work. Check them out. Not sponsored, just homies. He actually recreated an open-sourced beef. Do you remember beef? Oh, yeah, yeah, yeah. So I'm really excited to check it out. I was a big beef user way back early in my career.

Yeah. And beef is kind of like an abandoned project. If you guys don't know, browser exploitation framework, browser something. Yeah. I'm missing an E. What's beef? I don't know. I thought it was that too. Yeah, yeah, yeah. So the whole point of the project was you create these moment-in-time use exploits based on browser fingerprints and all this kind of stuff. It was a pen testing helper tool to exploit people's browsers. Really cool tool. Kind of went away or wasn't maintained.

And so Jason stepped up and was like, all right, we're going to maintain it and open source it and all this kind of stuff. So shout out to the two homies there. Hell yeah. Cool stuff. That's awesome. My other shout-out, our home on Tuesday for a bit, at least mine, was Decibel. I mean, they're a VC group doing cool things, invested in a lot of cool companies. They took over Libertine Social in the hallway at Black Hat, and it was just a really good—they invited us in. Endor Labs invited us in. They had, like, an influencer set up for us. Again, not sponsors. They were just really nice to us all week, so have to give them a shout.

And they brought— Honestly, some of the best and brightest that were at Black Hat spent a lot of time at that place that facilitated some really cool conversations for us. We ripped some Pokemon cards. We did, yeah. Endor, randomly. Again, a sponsor for me, but not for this podcast. They had Pokemon packs. They were just giving out people to rip. And so we ripped a couple. They were one Chaos Black and one Prismatic Evolution each. Garbage. Didn't get anything, but it was a good time. You get the dopamine rush of the pull. I have an Eevee that I got. They gave you an Eevee, but it was already pulled. It was a pre-pulled Eevee.

Yeah, not a rare one. I think it was like a mid-rare. I got nothing. I got bulk. One more shout-out we have is, before we get into the bugs of the podcast, the sponsor of this podcast, Maze. Good friend. Give him a shout-out. Maze, take it away. Hey hackers, before we keep going, I wanna take a quick break and say a quick thank you to our sponsors. Today's Lowdown episode is sponsored by our friends over at Maze. Maze has spent the last couple of years using AI to find and remediate vulnerabilities in the cloud. And now they've released Maze Code, AI agents that find and deeply investigate vulnerabilities in your code.

Now we all know the deal. Software composition analysis tools and SaaS toolings are all historically really bad at just making more noise. They find issues that may not actually be real and just create a lot of problems for your engineers to deal with that don't solve real security vulnerabilities. I get it. We've been there. Maze code is a really cool tool because Maze code actually knows how your code works. Maze code investigates every finding with the context of how your code is actually deployed. Sure, you're using a version of libxml2 that's vulnerable, but the actual code path to hit that vulnerability is not exposed in your Docker image.

Maze code can find that and won't show you that that vulnerability is a big deal. Every vulnerability that Maze code reports comes with evidence that it's real and a way to fix it. Check out Maze at go.lowdownpod.com slash Maze. That's go.lowdownpod.com slash Maze. Thanks again, Maze. Let's keep going. You're such a pro podcaster. And we're back. Welcome back, everybody, to the lowdown. Yeah, man, so cool. So what's going on in the cyber world, dude? What's happening here? I mean, that 1Password thing was the thing that I was really sticking on this week.

The other bit is, I think, really cool. OpenAI felt the need to show up last minute at Black Hat unscheduled and explain the hugging face incident from their side, right? Obviously, we're working off of our phones this week. We can probably go into it a little bit on the next episode. This talk just happened yesterday. But they were pretty transparent on stage with what they need. I've heard some criticisms. I've heard some kudos. So I'm excited for us to form our own opinion about it, but it's good to mention that they showed up.

I think kudos. They showed up. Yeah, 100%. Stood on stage, did it. This is kind of like last year. I was in the Innovators Summit on Tuesday at Black Hat, and... The leadership at CrowdStrike showed up, and it was right after they took down, like, all of the internet. It's a world for a minute, yeah. Yeah, and they showed up, and they got on stage, and they explained themselves. And I feel like OpenAI did very similar this week, so I say kudos for showing up and explaining it. Some people out there are, like, hoping for more transparency, like, especially actionable transparency.

what happened so that we can not let it happen while we're using your products or your tools, I think is what people are asking for. As someone who's led incident response, you're not going to get the audit logs. I'm never publishing logs. You never know. It's impossible to sanitize that stuff to that degree. I've seen people ask for that. I'm like, that's too much. Yeah, not going to happen. You know what I mean? I think the one reason that we'll ask is the prompt. Like, what did they say to the model to make it do the thing? Because maybe that kind of at least contextualizes what it thought it was supposed to do, like what kind of sandboxing it thought it had.

Like, just to set the stage for why it may have taken the actions that it did. Well, yeah, I mean, we were at dinner with Clint last night. He's at OpenAI. I mean, I don't think this is super secret. Like, one of their, like, as they test new models, they use very basic prompting to test the full capabilities, cyber capabilities. I think this is not super, super secret. But he's like, you know, it's a few lines, right? Just, like, literally hack this, don't make no mistakes is, like, one of their tests. I think he said, yeah, as, like, their benchmark of, like, danger, right? Like, if a 14-year-old got this model and said, like, find me a zero-day in Google Chrome,

Could it? And if the answer is yes, that's a problem. They acknowledge that a good model with a good harness will find the bugs. The question is, what is the barrier to entry to that process? That's the danger. I think it makes sense, right? Because if you look at some of the most successful zero-day finding things that we've seen post-Mythos era level stuff, it's not. go find this, make no mistakes. It is a well-oiled vulnerability research machine that just got steroids injected into it.

Those are not the same thing. Yeah, Yannet Blackhat said that. He said the new meta for VR is like, it's not that you tell the AI, find the bug, no mistakes. That does not work. It'll hallucinate bugs. It'll find some bugs, but not all of them. What you do is tell the AI, here are the tools at your disposal. Here is the unique problem. It's either you have the source code, you want to do white box fuzzing. You don't have the source code, closed source binary. here's the target, here are the tools, build me a system that uses the available tools for the problem, right? That is a much more difficult research question than find me zero to no mistakes.

And maybe the models will get there, maybe we'll improve, because we are seeing that kind of like a little bit of a hockey stick in the number of found, disclosed vulnerabilities, but I mean, hopefully we don't get to the true like, we found them all and we just have to deal with the consequences, you know? Yeah, absolutely. I'd say some of the other talks that I saw that I'm excited to, like, dig in a little bit deeper post this, like maybe reach out to some of the speakers. Maybe that's probably a good future content idea is, like, follow up with some people that dropped some cool stuff this week and get them on our channels. But Jamie over at Huntress gave a talk at Black Hat about, do you remember the story?

We probably covered it, about how one of the Black Hats, like— had their software on the machine and they did this whole investigation. There's a whole like Twitter outcry about like, oh my God, like Huntress is spying on us. And it's like, brother, do you know what EDR is like? Yeah. Yeah. Yeah. So like that as a talk, they kind of went through the whole investigation with that, like, and went behind the scenes and all this. I'm really excited to dig into that one. Yeah. This kind of stuff happens a lot. What was the, what was the recent one where I was like, oh, is this really something to worry about?

Or is this like when normies figure out what EDR is? Oh, it was, uh, the windows gd yeah yeah yeah they're like oh my god windows has telemetry yes i was like uh is this like when normies realized what edr is basically corporate spyware and of course it is yeah i mean we used to make the joke like uh when i was running blue team it's like oh i want to be the nsa for right like my thing your your feelings aside about government like patriot act level stuff right at your company you want that that is like you're on your team That's your job. That's your objective is to have a level. You need everything.

You need every command run. You need every website visited. You need to know everything that goes on on your corporate network. So, yeah, sometimes people are shocked that that happens. How about you? What kind of a bunch of research job that you've been focusing on? Yeah. So, I mean, on the topic of bugs just keep coming out, right, we keep seeing... This trend where new styles of bugs are being discovered through LLMs by one researcher, and then either that researcher or fallen researchers discover new variants of that bug. So in a previous podcast episode, we talked about Jan Escape, right?

This was a guest to host Virtual Machine Escape. in a nested VM, right? So in the world of nested hypervisors and nested virtualization, you are the VM host on what's called bare metal, and then you host another hypervisor that maybe you're renting out, and then someone hosts their VM inside of that. So what you have is a structure, which is an L0, the native host, the L1, which is the virtual hypervisor, and then the L2, the second layered guest. The vulnerability was a use after free in the way that the shadow MMU, which is basically a structure for tracking memory pages of the guest, got handled in the kernel in KVM, right?

This is a Linux hypervisor. And so this researcher in finding Januscape... got paid out 250 grand by KVM CTF, because that is a CTF put on by Google, where if you can escape from a KVM VM on Intel hardware, they will pay you 250 grand. So if you're interested in paying off part of your mortgage in the next couple of months, go please participate. And so that same guy, the same god mode of one researcher, Hi-On-Woo Kim, if I fucked your name, man, I'm really sorry, found literally,

Same class of bug, use after free and shadow memory pages in an L1 virtual machine. So just not even nested, just like you're in a VM on KVM using that same structure of code. Right. We're on AMD hardware, which is called SVM for virtual machines. The shadow MMU bookkeeping that can lead to use after free causing a guest triggered page fault and potentially use after free enabling remote code execution. Right. So it's just interesting to see these really unique classes of bugs. Also, on Twitter, I saw someone leaking that in the future they're going to drop another bad pipe, a pipe failed site vulnerability, which is basically like, you know, in the kernel when you pipe two files together, you have the pages that they live on in page cache.

And if you do operations on them temporarily, it might overwrite pages that should not be overwritten. And if you use that page cache later, you can do things like overwrite the current page cache entry for like a set UID binary. So like run that code as root, right? Really interesting new styles of bugs that once they come out through some kind of LLM research, there are a hundred different implementations of that bug in the kernel that get popped. So a lot of those popping out right now. I feel like it's going to take us three weeks to cover all the research that dropped. Dude, I've been at DEF CON for four or five days, and I feel like I missed like 15 bugs to talk about.

I'll bring up one. I made a video about it. We can probably link that goes deeper into it, but we can talk briefly. Another Shai Haloud supply chain attack hit. Literally, the threat actor said, Shai Haloud, here we go again is one of the IOCs. It's like in the GitHub repo, all this kind of stuff. massive, massive packages that actually got compromised. This Kivi was the main one, but all of the packages, flat cache, file entry cache, we're talking 2 billion downloads a month on these packages.

So like crazy thing. And for context, I think Key V on its own was like $174 million. $604 million a month. Flat cash, $580 million a month. I mean, just crazy, crazy, crazy. So just if you guys don't remember Shai Halued and why this is so crazy, what happens is this malicious code gets into these JavaScript packages. And... Hacks the actual developer who has NPM credentials to publish the packages. And generally these people don't just publish one package. They're prolific JavaScript developers.

And so they can, they have access to repos for more packages. And then it just becomes a self replicating worm. It just gets the maintainers of that package and that package and that package. The thing that like really stands out to me, we've talked about Shia Lude. I know I've talked about it on my channel. The thing that stands out is the threat actors are stealing lots of other information as part of these attacks, not just NPM tokens and GitHub tokens. It seems like most of what we see in terms of impact is just self-replication, which is just NPM and GitHub token usage, but they're stealing...

your HashiCorp vault, your AWS creds, your Kubernetes creds. I'm looking at the list. I mean, Stripe stuff, generic environment variables, private keys, SSH keys. A lot of AI credentials too. Terraform stuff, AI stuff, VPN stuff, key pass stuff. Where is the ripple effects on all this? Right. It almost makes you wonder if they're just purely profit incentives, and they're like, they're selling them in the black market, right? Like, just, let's get some cash and get out. I haven't even heard that. Because there are rumors that they're, like, teenagers, right? They're, like, young people, isn't that? Well, Team PCP. I don't know if Team PCP is behind every iteration of Shia Haleed.

Yes, they were part of it. Yes, I think they're kind of YOLOing, very YOLO threat actors. But still... If you've taken the time to code all of this other secret stealing code into your malware, it keeps me up at night going, what else got stolen that hasn't been used yet that we haven't seen a second order effect of one of these NPM works? I do understand. Yeah, that is kind of terrifying. Yeah, that's one of the keep me up at night things. The other thing, kind of a follow up from one of our main stories last week was the water supply hacks that we obviously talked about were attributed to Iran.

The government is like at this point screaming about this because those PLCs that you talked about in the last episode are still on the Internet and still confirmed vulnerable to CVEs from 2013. And like, what else do we need to do to like get these things off the Internet? I don't understand. Yeah, so when I went to the General Nakasone press briefing earlier, that was one of the conversations too. It was like, what does the government have to do to get these things off the Internet? And it's like, guys, that is just like, there should be a program, your company that does the awareness training that's like, hey, maybe we should figure out if our PLCs are on the Internet, at least for the defense industrial base.

He said that the DOD offers a service that they will just scan your stuff. They'll be like, okay, let's go look at your website, your IP address range, and see what ports are open to the Internet. is it a PLC? Maybe it shouldn't be. I don't know. I would hope or imagine that at least the DHS for, like, private U.S. companies would do something similar. But, yeah, that's a whole other conversation about, like, the defunding of CISA and, like, the irony of, like, Colonial Pipeline, new administration, CISA defunded, Iran hacks Minnesota, right? So it's a very depressing story.

Yeah, it's wild. We did get to meet briefly that one of the old heads of CISA, something or other, that we met in the hallway. It was a CIO, I think, or the CISA. Yeah. that might want to come talk to us about this kind of thing, because I would love the insight, right? To me, it seems like people with these remits should be able to pick up a bat phone and say, hey, that's vulnerable to a 20-plus-year-old CVE. It's on the internet. Where's the bat phone that says, like, pull that off the internet? In the press briefing, he said, I didn't realize this, in the U.S., there are 50,000 water municipalities that all have to get maintained.

And it's like there's multiple problems there. One, every one of them is subject to some form of state funding that they don't get for security. And then on the federal side, A, the federal government doesn't really have the authority to regulate that. And even if they did, how are they going to regulate 50,000 water plants, water municipalities? It's this very... As anything is with security, right? It's a very complex scaling problem that no one has the answer to. Yeah, I guess I'm probably oversimplifying it in my head because I've worked in big megacorps where I had that bat phone, so I'm kind of expecting that, right? If I found a vuln at, you know, large bank, insert large bank that I worked at, like, I had the power to, like, call and give that person a really hard time until that thing disappeared, right?

Sure, sure. I just feel like even though they are 50,000 different water municipalities, their critical infrastructure... they're beholden to federal funding and regulation, there should be someone that has the, I agree, like political clout to call someone at these places. And it's like, what are you doing? Like the water supply is like what we're talking about here. in that scenario, people know that it's internet connected. I don't think anybody knows. I think it's one of those scenarios where there's just literally a lack of knowledge on where the borders start and stop.

You know what I mean? Which again, for a mom and pop, five person company or whatever that's maintaining the pump in rural Kentucky, I don't know. I just don't know if they have the funding to maintain a team that actually has that knowledge. We know this from other incidents. So much of security is just like a knowledge management failure. Like, oh, Jim was implementing that firewall rule, and then Jim went on vacation, and then Jim quit. And no one wrote down what Jim was doing, so it never got finished, and then we got popped because that hole was still there.

You know what it is? One of the hardest parts of getting vulns fixed in large organizations, and this isn't even a large organization because there isn't like an umbrella thing unless you talk about the US government as a whole, is figuring out who to call. Sure, yes. A lot of times you see this phone, you see this thing, you're like, who the hell owns this? I had an entire working group on my team that had people throughout the org, through devs, through engineering organizations that was called the code owners working group.

And it was just like, figure out... who's the code owner for all these repos, for all these things, because if I found a vuln and a repo, it wasn't a given that I would know exactly who to call to get that thing fixed, or open a PR and get the PR merged, or whatever it was. So I can imagine this is that problem on dial turn to 11 right when you're talking about all these different municipalities yeah organizations involved for sure makes it complicated all right man what what what else at vegas are you super stoked about before we wrap up hmm what else at vegas i'm going to a party with some of the flair homies later this one tonight that'll be good

I'm going to spend some time tomorrow at one of the CTFs and try to hack a robot. Maybe win some cash monies on that. It'll be kind of fun. And it's going to bop around the DDoS village, the IoT village, and just see what fun firmware I can rip apart. What about you? Yeah, same. We've done a few laps around here. I really like going deep into these little corners in the villages here. We've got some homies to go say hi to at some of the booths, the Hack5 family over there, the Miscreants family. They were kind of swamped when we went over there earlier, so I'm excited to go see them.

Yeah, man. Just, like, feeling the passion. I think you said this a few weeks ago, that you come out of Vegas a lot of times energized, but, like, full of ideas. We've been bouncing ideas off of each other about how we can kind of continue to level this up. I gotta go get some of that. I gotta go find some of that passion from some of these village organizers and just mainline it, and be like, yeah, gimme, gimme. Yeah. I always walk away from DEF CON Black Hat physically exhausted, but like spiritually rejuvenated.

Yeah, yeah, yeah. You know what I mean? Yeah. I think I'm glad we did this because otherwise we were going to have to record a podcast when we got home real quick, and I do not want to do anything when I get home. All right, y'all. Well, thank you for watching as usual. If you're new here, hit that sub button. If you're on Spotify, give us a rating. Five stars preferred. But if that's not how you feel, then it is what it is. And yeah, man, again, we recommend DEFCON and Black Hat very highly. If you haven't done it, you should do it. And Mays, thank you again for partnering with us and keeping the world safe. Yeah, that's it. Thanks so much. Wrap it up.

See you guys.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Java4Shell, NGINX2Root & the Week Everything Went Down

    Nov 20, 202658 minAudio

    Recorded Wednesday, November 18, 2026. Java and NGINX had a bad week. So did anyone trying to catch a flight. The Low Down Podcast is presented by Maze. https://mazehq.com/lowdown X: https://twitter.com/Maze_Security Ed...

    Credit the people on this episode

  2. This SSH Bug Accepted Wrong Passwords, AI Exploits, Tiny Keys & a Surprise Guest

    Nov 17, 20261 hr 11 minAudio

    Recorded on Friday, November 13, 2026 The Low Down Podcast is presented by Maze. https://mazehq.com/lowdown X: https://twitter.com/Maze_Security 257 A's should not get you into an SSH account. John Hammond joins Ed and...

    Credit the people on this episode

  3. Binance's Fake Interview Hack, Railway's AI Incident, Hacked Freezers & a Critical Linux Bug

    Nov 6, 20261 hr 6 minAudio

    The Low Down Podcast is presented by Maze. https://mazehq.com/lowdown X: https://twitter.com/Maze_Security The "signup test" was probing for Kubernetes secrets. Ed and Matt get into Railway's AI testing incident, hacked...

    Credit the people on this episode

  4. Black October Recap, OpenCode’s Stolen API Keys, AI Distillation & the LiteLLM Supply-Chain Attack

    Nov 3, 20261 hr 24 minAudio

    Welcome to The Low Down, the best show on the internet for hackers Black October has real bugs, stolen keys, and a lot of very loud headlines. Ed and Matt sort through PHP2Shell, AI security claims, and the difference...

    Credit the people on this episode

  5. Cybersecurity Experts REACT to the php chaos... (and MORE)

    Oct 28, 20261 hr 28 minAudio

    Welcome to The Low Down, the best show on the internet for hackers Listen on Spotify! https://go.lowdownpod.com/spotify The Low Down is presented by Maze. LinkedIn: https://www.linkedin.com/company/mazehq/ X:...

    Credit the people on this episode