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

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

Nov 17, 20261 hr 11 minAudio
ShareXLinkedInFacebookWhatsAppEmail

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 Matt to talk about N-able, ScreenConnect, and a rough week for remote support. Then we get into the SSH login bug, OpenAI's Linux kernel research, FortiClient, and eMudhra's tiny private keys. There's also some World of Warcraft, because apparently we needed a break. Today we're talking about: John Hammond, N-able, and ScreenConnect John walks through the RMM mess, including SCKit and the difference between a vulnerable server extension and a vulnerable technician client. We talk about how remote support can spread an infection, then get into incident-response burnout. N-able report: https://www.huntress.com/blog/n-able-disables-cloud-services-following-disclosure-of-able2shell-exploit-chains SCKit research: https://www.huntress.com/blog/sckit-worm-rogue-screenconnect Client advisory: https://www.connectwise.com/company/trust/security-bulletins/2026-09-08-screenconnect-bulletin Exploitation in the wild and local AI Why doesn't every scary bug turn into an internet-wide fire? We talk about exploit conditions, GreyNoise's PHP observations, a separate AI-assisted intrusion report, and running local models for security research. GreyNoise PHP tracking: https://php.events.greynoise.io/ Operation Snowball: https://malbearlabs.com/posts/operation-snowball-inside-an-autonomous-ai-intrusion-campaign The SSH login bug Ed explains how OpenSSH and musl can turn a password longer than 256 bytes into a successful login on affected configurations. We follow the asterisks, the old fixes, and the practical problem of figuring out which devices are exposed. Original write-up and comic: https://sshhole.com/ OpenAI's kernel bugs and Daybreak OpenAI reports 948 working local privilege escalation exploits. We unpack what that number means, the API-credit cost, and the research pipeline. Then we ask what happens to AI security vendors when the model companies start offering the tools themselves. Linux kernel research: https://openai.com/index/a-thousand-fewer-paths-to-root/ Daybreak Red Team: https://openai.com/daybreak/red-team/ Daybreak Defense Factory: https://openai.com/daybreak/defense-factory/ FortiClient and the security-vendor problem A FortiClient/FortiPAM exploit demo leads into a bigger argument about edge devices, old toolchains, and why preventable bugs keep shipping in security products. Researcher's demo: https://www.linkedin.com/posts/kevin-joensen_cve-2026-121529-or-fortinet2shell-as-we-ugcPost-7526030522238697472-3JP7/ eMudhra's weak private keys One private key was 4. We talk about the HSM migration mistake behind eMudhra's weak intermediate-certificate keys, what the hardware was supposed to protect, and why checking the result matters. Company incident report: https://bugzilla.mozilla.org/show_bug.cgi?id=2083449#c82 World of Warcraft and logging off Alliance or Horde, busy-dad gaming, and finding time to play something without turning it into another job. Which is harder to keep up with right now: the vulnerabilities, the patches, or the AI product launches? Linked reports may include later updates. ---------- 0:00 Intro and John Hammond 3:59 N-able and ScreenConnect attacks 13:48 Exploitation in the wild 17:28 Jailbroken AI and local models 22:48 The 257-character SSH login bug 32:42 OpenAI’s Linux kernel exploit research 44:03 Daybreak and AI security vendors 49:55 FortiClient and edge security 58:14 eMudhra and the private key of 4 1:04:13 World of Warcraft ---------- Listen on Spotify! https://go.lowdownpod.com/spotify 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 Ed / Low Level Instagram: https://www.instagram.com/lowleveltv/ LinkedIn: https://www.linkedin.com/in/lowleveltv/ YouTube: https://www.youtube.com/@LowLevelTV Matt Johansen / Vulnerable U Instagram: https://www.instagram.com/mattjayy/ LinkedIn: https://www.linkedin.com/in/matthewjohansen/ YouTube: https://www.youtube.com/@VulnerableU John Hammond YouTube: https://www.youtube.com/@_JohnHammond

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.

And it's not a memory corruption issue. It's not a race condition. It's not a business logic issue. I guess technically it kind of is. It's literally, we forgot that these two functions did a certain thing. And now when they do the correct thing, it means you logged in. That's the bug. Insane vulnerability. Insane thing. Good morning, hackers. Welcome back to the best podcast on the internet for hackers, The Lowdown, presented by Mays.

Matt, I got a bone to pick with you, bro. Okay. You were late to the podcast. Did you not get my email? I didn't get any email. Is there a particular reason why you would have gotten no emails today? More like Microsoft 364, am I right? Am I right? Microsoft 503. More like Microslop. Oh, my God. Yeah, bro. So we're filming this episode on November 13th, which at this point, I think Microsoft Outlook has been down for over 24 hours, like an entire day of global downtime.

Not one company, not one region, global Outlook outage. Now, a little strip tease there. If you expect us to talk about that, we're not going to because no one knows any details about it. But that's just kind of hilarious that it's happening right now. Dude, I was, again, like, I measure things in our industry by how many, like, normal people in my life start reaching out to me about stuff. I've gotten a bunch of text messages this morning from people like, what is going on? That's so funny, dude. From, like, random other, like, dads, like, friends of my kids, like, stuff, you know.

I figured I'd go close to the source. You know, Matt's probably got to know what's going on here. And I was like, I actually don't. I'm in a bunch of, like, sports. speculative text messages groups with even people that like should be in the know. And I don't know if we're just like not hearing stuff. I'm guessing this doesn't smell like a cybery thing to me. I agree. Whoa. Who's that Pokemon? It's just... A wild John Hammond has appeared. What's up, John? How you doing, man? Oh, hey. Hey, everybody.

How are you, brother? Goodness gracious. I'm so sorry for jumping in. It's all good. I know, I've heard you may be busy with a particular Enable thing going on. Is that correct, or is that something else? Yeah, a little bit of Enable, a little bit of Screen Connect, a little bit of, you know, just the internet doing what it does. Yeah, but before we get to that, we're talking about the Microsoft outage, which I'm sure you may be affected by minorly at least. How are you aware of what the hell is going on over there? It sure sounds like just a big outage. Yeah, I think there was a latest thing that said like, yeah, some CPU just keeps falling over or something.

Yeah, good sign. Love that. Yeah, it smelled like DNS to me, right? It didn't smell like cyber. It smelled like something sillier, right? Cool, man. Well, thanks for joining The Lowdown, you know, the best podcast for hackers on the planet, as we like to say. This is quite an honor to be hanging out with you all. If I may say, I do have another call in like 25-ish minutes, but I would love to hang out if you're willing to have me. Go have the footage, edit, whatever you need. Yeah, we're just yapping. We're opening the show. We're yapping. Matt was talking about how he grades the severity of an incident by, like, how much he gets asked in real life from his other friends about, like, that are outside of the industry, what they say.

I was going to say, I haven't been getting a lot about Microsoft, but I keep getting questions about, like, have you heard that the grid is going to go down? I'm like, what? And they're like, yeah, like, world leaders keep saying we've got to stockpile food and water. I'm like, I have not heard that, but I'll keep an ear out. Have you guys gotten any of that? Yes. Yes. I know of some people going like full-on survivalist mode in their spook and concern of, yeah, the world changing. Yeah, dude. It's weird. I mean, AI is good at security and hacking.

And as we've kind of seen, the OT community most of the time, sorry, guys, is not. So we'll see how that plays out. Yeah. It's not great. Yeah. So if you guys don't know John Hammond for some reason, I don't know why you wouldn't know John Hammond. Between the three of us, we have millions of YouTube subscribers, by the way. Wait a second. Yeah, but John, you, I mean, Ed and I, this is our job now. You're looking at it. My job is beach. My job is coffee. My job is podcasts, right?

Very Ken in the Barbie movie of me. You have like an actual job. And your actual job is talking a lot about like one of our main stories of today, which is this Screen Connect thing and Enable thing. I was reading a bit on the Huntress website about both of them. seems like a really bad day to be an RMM world, right? Or an MSP. Can you just give us the download of like what the heck is going on with Screen Connect and Enable?

I can do my best. I feel like I personally have been tracking a lot of the Screen Connect and Enable more than I have Enable, but I can try to go to either one of them. But the gist is, yeah, it's been a pretty rough week for our managed service providers of the world. Let me do my best with Enable because I can, I don't know, maybe try to answer some of that and then we'll go through it. But an individual out on the internet had just dumped a whole lot of proof of concepts for a whole lot of vulnerabilities for Enable.

And you've heard that sweet, like two shell suffix that we're just adding to everything now. We gotta stop. We gotta stop. Yeah. Uh, and they call it able to shell and or agent to shell. Now there's forever able to shell because he's added more. Um, but Enable is aware, obviously they took a heck of a lot of things offline so that that wouldn't be as susceptible as it is. Um, and then, uh, And that has been, I think, from disgruntlement of the researcher not hearing back from Enable, but I don't want to be like pointing fingers or casting shade in any of those directions because I think there is a more thorough process for an actual report to Enable, not just an email and hope someone responds in your inbox.

I don't know. I wish I knew. But- What else are you all worried on the edible side? Well, I had to, first of all, you know, new noob style here, the I'm not going to do myself any favor in the YouTube comments that called me stupid over the last few weeks, but I had to figure out what the heck edible was, right? I wasn't familiar with that to begin with. I haven't run into that in a bit, but it seemed to me, I mean, the POC itself was kind of wild, right? It's like go to website calc exe pops. It's like, okay, that's, yeah, that's, that's about as POC as you get.

We have another one, right? Fortinet did the same thing this week too. So like, and then I was even looking at our friends over at Grey Noise and they're starting to see a ton of activity actually scanning for this edible thing as well and active exploitation start to go down. Yeah, I mean, it just seems like a wild one that we all got to get our head around and like, you know, this enable or end central users just need to watch this, watch this active exploitation campaign going on. Forever.

Oh, number four ever of both who shall. So I'm so over the to shell, man. It's like dirty something and something to shell. Like, please stop. Just let's get some new words out there, please. Yeah. We're seeing at least one successful intrusion according to gray noise that I'm seeing right now too. So yeah. Yeah, seems like a spicy one. But okay, you know more about Screen Connect though. Yeah, I've been chasing a little bit of the Screen Connect side of the house because that one is wild and has a lot of different kind of interesting things to chase.

What we had shared previously over on the blog, if you are willing to take a look at the old Huntress write-up. Yeah, I've got it up. There's this thing called SC Kit, the, I'm assuming Screen Connect Kit. And that is a worm so that, okay, when a user or technician would connect into a screen connect, like usual tech support, the good intentions of remote monitoring to managements, it would propagate to other affected hosts. But the way that that one came in through the door, the patient zero of this intrusion is a little bit more spicy because it's server side.

And there's a lot of nuance on the client side versus the server side for screen connecting because – This actually has some of the spooky little supply chain shenanigans to it where it's a plug-in or extension for Screen Connect server that has some weak vulnerability in itself. This is RMM+, that would just be an open door to anyone being able to point at the Screen Connect server that's running it and then have the capability to push downstream to clients and have that worm-a-gate. Yeah.

This is off the tails of some of the previous Screen Connect worms that have been in the news and the headlines lately that we were chasing previously. And I think the gimmick to that was a CVE. So specific CVE number, CVE 2026-84869. That's burned into my brain for the past couple of days. That is on the client side. That is when the technician were to connect to another screen connect computer via RMM capability, it poisons them.

The technician. So yeah, yeah. Their computer, the connecting guest would have code execution onto the technician's computer. So that's how I believe a lot of screen connect kit or SC kit has kind of propagated more. Um, and that's what I think a lot of the community is now trying to get the nuance around. Okay. The version number for the server isn't as much as a concern, but if they're patching and updating the clients that their technicians use, that has to be at this version, 26.6.5, which is now fixed. But yeah. Now there's a new strange thing that I'm going to have a blog out for today as quick as I can for other people screaming their head off on Reddit of weird cryptocurrency connections.

With screen, having to do with this? Having to do with screen connection? Yeah, yeah, yeah. Here, I can drop a Reddit link for you and maybe I can give you at least some of the inside breaking news. Yeah. It's out the door soon. So let me get this straight. So Threadactor wants to Wormify via RMM software. They get a technician to connect to them, which then infects the technician, which then sets off this cool graphic that you guys put of 288 hosts in like minutes.

Because that's what that technician then had access to. Do I got it right? Right. Yeah. Wow. Okay. That's a really, really cool one. Just so I understand as well, this originates from a vulnerability in a plugin that the technician has on their computer. The server has. Right. Okay. Okay. The screen connects server has it. The technician using the server to reach their other hosts would infect the first round of technicians and the technicians connecting to other computers continue to infect and infect and infect and infect. So is the threat, do we, and maybe I'm getting to a point where you're not allowed to say, but is it like the threat actor is getting the technician to connect to the computer or is the, is the computer infected by malware initially that then is, you know, doing this?

You know what I mean? Right. I'm the former when you get to the client side. Um, and I think, I hope that's what some of the new stuff that I had pointed at to you will at least have a semblance of sense to it because now I think there is a running insight of, look, could we bait people? Could we lure and have some social engineering to have them click on the computer that will then infect them? Um, so that should be a fun story to tell as quick as I can.

Yeah, it's wild, man. Yeah, you're well, thank you in the middle of firefighting phone calls, hopping on the lowdown. I know not on the highest of priorities, but we appreciate you. You guys are cool. I want to hang out with you all. And forgive me if I'm like too frazzled or frantic, if I'm not explaining things all that well. No, you're great, man. Dude, so... It's the end of the year, and a couple months of, like, it feels like back-to-back-to-back kind of insane vulnerabilities. So, yeah, I think we're all just, like, mentally spent, dude.

We were, like, put the show notes together for today, and we only had, like, two CDSS 10-crit, like, volumes in major software. We're like, wow, slow week. Like, you know, this is kind of crazy. We might go short on this podcast today. Yeah, the other thing, man, I've been trying to figure out how to make some content about this. and we should all talk about it, is I've been talking to people in the incident response side of things. I stay in touch a lot with Threat Intel groups. I know you guys are in some of the same groups I am.

And then talking to some incident responders, that that's what they do for a living. They are so far beyond maxed in terms of what they're able to do right now. And they are trying so desperately to hire people So if you're like struggling in the job market right now and you have any sort of chops to like get into incident response, they are hurting, hurting with the amount of work that they're getting called into, right? Because of all this stuff, we all get to chuckle about the headlines, but it turns out that like...

When we see active exploitation and that's it, we say, oh, yeah, this is being actively exploited. Turns out that means there's a lot of teams out there that are literally doing these firefighting drills, calling in the lawyers, calling in the incident responders. And literally law firms, too, are completely maxed on, like, ability to even take on this kind of work. It's kind of insane. Very insane. Tough to be a cyber guy these couple days. Cyber person, sorry. Well, one of the other, I know, I know you got to hop off soon.

One of the other stories that we were going to talk about, John, was the SSH bug, which feels like an eternity ago, but it actually came out since the last time we recorded the podcast. And we can get into the actual details after you hop off. But one of the things I would love to ask you is like, why some of these bugs like SSH one, the PHP one, like... I'm really interested in your take on like why we're not seeing the like absolute internet fire that like these things could have turned into. Right.

It's like very widespread, questionably, easily exploited. Maybe that's part of the answer is like, you know, the exploitation is kind of hard. But why do you think like PHP and this SSH bug aren't like absolutely wildfires right now? Yeah. I'd love to run down the road on the SSH to shell, but I might need your hand holding because I think the gimmick of that was just, uh, 255, 256 or like longer password would just let you write into root on like Alpine Linux and things.

Is that right? Yeah. Yeah. So many people have been poking fun. Oh, Terminal Tuchel. My buddy made a really good one. Pseudo Tuchel. Yeah, I saw that. It's so good, dude. It's so funny. Anyway, to your point, I probably would have a better answer on the PHP Tuchel side of it, not being as cognizant of what's been going on running around for SSH Tuchel. PHP to Shell specifically, which was one of those like, oh my God, the world is falling, the internet's on fire, everyone everywhere could be affected, vulnerable, exploited, why isn't there RCE, absolute Armageddon apocalypse.

That one was like genuinely at least very, very nuanced to the environment, to the version, to the heap grooming, memory corruption, low level, low level shit. that I don't know if, Ed, you might have some sweet insight on and all. Sorry, we have an inside joke. Whenever someone says low level, I point to the sign. It's become like a tick in my videos. Whenever I say the word low level, I say name of the channel, by the way. Yeah, that's it. That's it. Yeah, man. So, yeah, my understanding of that is my understanding of the reason why it wasn't a super hot, like, actual threat in the wild is for that exact reason, right?

The amount of recon you have to do that I think technically isn't really possible to know not only Apache or NGINX, What is the configuration of those two things? Are you reverse proxying to a different server? Are you on the same box? And then what version of PHP? And then from there, pulling off a race condition via a session establishment. Like, all those things have to line up perfectly to get the shell. So, like, I would imagine some nation-state, some highly-resourced threat actor could probably pull it off and probably has. But for the...

Mass exploitation scenario, I think just statistically, it's very difficult to get correct. So that's probably why. But I think following on to that, though, the SSH one is not that. It's literally a logic bug. So I'm shocked we haven't seen more of those as well. Have you all seen the Gray Noise page for PHP to Shell? Actually, I haven't looked at that. No. Oh. I think I've been focusing on the census one a bit, on the patching. Yeah, census one is great too. Big love to GreyNoise though. They had ones that saw confirmed exploitation, and you can see a lot of the tradecraft, and a good many of those hosts are pulling down payloads from what was an open directory.

So we had some fun shenanigans trying to see, oh, whoa, yeah, there is an exploit here. There are their victim data, and we got to do some sweet reporting and try to help clean this up. Um, but Gray noise, I know has been tracking as well as census and others, but I'm right there with you on the SSH to shell one. How come that hasn't been sprayed and prayed as much? Yeah. Kind of weird. Right. Um, the, The other was that open directory thing. There was a story that didn't make our show today wound up on the cutting room floor here.

But there was a story where a threat actor directory was found open and they found like a jailbroken deep seek. in there. Did you read about this? I didn't, but that makes complete sense. Yeah, they found a jailbroken deep seek in there and then they found evidence of this person being in an unnamed European telecom company. who hadn't announced that they were compromised, but this directory was like how they found it.

And so, but then like, as the, as the research dug deeper, and maybe now like we're actually talking about it, we can, we can pull some of it up. But as they dug deeper, all the vulnerabilities that were used in all the hacks were like old, crusty, you know, old enough to vote CVEs. You know what I mean? Like really old, boring CVEs that were actually used. So like the whole jailbroken deep sea thing was like, cool, didn't need, you know, the obliterated zero day machine to get into this European telecom. Right.

Oh, this is a Malbear labs. This is Anna. This is her work. Excellent. Yes. That's the one that's, that's unrelated to the other open directory from the threat actor that you were just talking about. Let me ask you all, do you all run obliterated models? Do you have any in your back pocket? I do not. The only model that I've run, I actually haven't even ran a model locally. I use like Olama hosted GLM-53, but I have not used an obliterated one, no. I haven't had the need to. Yeah. I have gotten my QN 3.8 obliterated up and running, and it...

It is obliterated, you know? It flaps. It flaps. You ask all the questions that you shouldn't ask, and it gives you the answers. What's your stack? Are you running it locally on like a GPU farm or what? Yeah, I have one beefy machine that can probably just fit like one or two maybe good running threads of 27 billion parameters. But I do have some friends in high places who have helped give me extreme compute power. There you go. So I'm running with the robots. I know I've told you all before. I'm a cyborg at this point. Oh, I'm with you, dude. I used to be, like, very anti-AI, and I think just seeing how good it is at VR, I'm like, oh, bro, if I can get one running in my house, I'm fucking going to.

They are too good to ignore at this point, so I'm with you for sure. Yeah, I'm boring. I'm doing mostly coding stuff, less VR these days. So I haven't had like the need to dabble. I've been trying to do some more of the local or the open weight model stuff. I don't have the sparks on my desk or anything like that. I'm going to need to bully you to see if your friends in high places are also my friends in high places. We do run in similar circles. But yeah, no, I'm like a frontier model fanboy at this point.

So... I mean, they're doing a great job too, but once we start to get to the super security security questions, they... Oh, no, totally. I mean, we were just talking off cam a bit about a project that Lola was working on. It's like you say the word cyber and you're in Opus 4.8 jail, right? Yeah. Yeah, I literally, like, I was writing a tool to, like, visualize firmware. I'm like, hey, here's the firmware. Show me, like, how the different processes talk. Is it network? Is it IPC? I just want to know. Draw me a diagram for security research. And it's like, ah, that's cyber.

Can't do that. I'm like, okay, man. I said firmware. Sorry. My bad. So, whatever. The F word. Yeah, right. Exactly. Exactly. How dare you say firmware? Yeah. uh yeah dude well cool i appreciate you come giving us the screen connect download we'll give you uh we'll give you some of your firefighting day back and uh why don't we have you on for like yeah why don't we have you on like with with less of a short notice and like you come on and we could talk about some some cool topic that we're all making videos about some week or something absolutely good to see you all have fun john hammond always down for a three-way see you later buddy take care bye

Yeah, let's go. Okay, cool. Hey, hackers, before we keep going, I want to 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 lib.xml2 that's vulnerable, but the actual code path to hit that vulnerability is not exposed in your Docker image. MazeCode can find that and won't show you that that vulnerability is a big deal. Every vulnerability that MazeCode reports comes with evidence that it's real and a way to fix it. Check out Maze at go.thelowdownpod.com slash Maze.

That's go.thelowdownpod.com slash Maze. Thanks again, Maze. Let's keep going. Okay, and we're back. Um, yeah, so we talked a lot about that SSH bug there. You want to actually go into what we're talking about? Please, please, please. Okay. Uh, again, uh, you know, I know I'm your cohost. I'm also, uh, just a fan of your channel. You did a really good video on this. You did, you did like, because I read all about this bug on Twitter and I was still kind of like, what, like, what, why? Like the joke I made was the fact that

Three different times this year, you could have tried between one and 257 the letter A's as like, you know, to break a secure protocol. And it would have worked three times in the last couple months. It's just insane, right? Yeah. So, you know, we talked a lot about the HSM and the key generation ceremony for the last one. This has nothing to do with anything like that. What was the root cause under this SSH bug? Yeah, so when I was reading about this on Twitter, you know, you see the POC, the guy demonstrating, and he puts literally 257 A's into an SSH login field for root, and it works.

And so your initial thought there is, oh my god, buffer overflow, is it 1991 all over again? That is actually not what happened here. And so this is a really, really interesting vulnerability that is a function of what I call contract drift, So you have two functions that you kind of assume one does one thing, one does the other. And as time goes on, you kind of forget how they work, and they go farther and farther apart until you get to a point where you forget that both of them depend on each other. You forget exactly how, and that forgetfulness creates a vulnerability.

Okay, so I'm literally going to read from the comic that the author of the bug, the finder of the bug, made to talk about it in simpler terms. They have a whole write-up, and they summarize the write-up in this comic. Okay. Okay. This comic was great. Like, honestly, like for complex topics, this was super, super helpful. So basically back in 1980, right, when they're making the whole like password system for Linux, Bill Joy in the Etsy password schema says, hey, we need a character or a symbol to represent an account that does not have a password.

Okay, in this case, it was the daemon or the daemon account that should not be able to be logged into. And so he says, okay, cool. The star, the asterisk, right, shift eight, is going to represent does not have a password. Okay, not a big deal. Fast forward 42 years, August 2012. So if you're not aware, When you're writing in C, a lot of the code that you write is not actually code you've written. You use functions like printf, use functions like read or whatever. Those functions live in the libc standard library. Now, there are multiple distributions or versions of this library.

The common one is GNU libc, glibc. Another one is called MUSL. I say Moosl. People have been yelling at me that it's Muscle. Fuck you guys, it's Moosl. There are different variants, and this bug is only in the Moosl variant. But basically, there's a function called crypt that does the password hashing and creates the password hash string that you see in your now Etsy shadow files, okay? For a long time, up until 2012... That thing, if the crypt function failed, would return a null. It would turn literally just a null pointer. The problem was programmers are bad sometimes.

Don't check the return values. And because crypt could fail, it would return null. You would dereference null would crash your program. So they said, okay, we need to have a better way of returning failure. And so instead of returning null as failure, we're going to return something special. We're going to turn something real cool, real neat that no one's ever heard before. It's also the asterisk. It's also the same character that Bill Joy said is no password. Not a big deal. Surely that won't bite us in the ass. Fast forward four more years, 2016. There is a timing attack that is disclosed in SSH.

So let's say we're on a computer, right? There is an account for Bob. There is not an account for Matt. OK, if I want to enumerate the accounts on the box, what I can do is log into a known account or sorry, I can log into an account that does not exist like Matt. And I can time that it immediately comes back because there is no password. There's no hashing done. The CPU does not run for a long time. I can log in to Bob's account, or at least try to, and see how long that takes. I can use that timing to figure out what other accounts do and don't exist.

So to fix this, they made it so for accounts that do not exist, you use the algorithm of some other random account, right? This will matter here in a second. Um, now remember how we said before, if the crypt function fails, it returns star. Okay. Well, what happens is if you try to log into an account that has no password and the crypt algorithm fails, it will return a star. And that star means no password. So when it compares to two hashes, that passes the login algorithm.

And it turns out for SHA-512, if you give it more than 256 characters, it will fail out to prevent a DDoS. Because if I gave you 48 million characters to process, it would spin the CPU. It would occupy the CPU. So I said, okay, we got to have a max out. We can't have too many characters. So what you do is for a SHA-512 account, You give it 256 characters. That fails out to zero in the child function. The parent function returns a star, and then that star matches the star in the EdSite shoutout file.

That's the bug. Insane. And it's not a memory corruption issue. It's not a race condition. It's not a, like, business logic issue, I guess, technically it kind of is. It's literally, we forgot that these two functions did a certain thing, and now when they do the correct thing, it means you logged in. insane vulnerability it's wild all right so ssh not that widely used right no definitely not so to be clear this is only the open ssh server distribution in muscle c now this is why alpine was kind of the talked about hotness because the alpine account is i think um no login by default no no what is it it's like

Yep, okay, if you log in as Alpine via this way, you get in as Alpine, but Alpine is the Alpine pseudo-equivalent, so you're basically root, effectively. Alpine is vulnerable to this. OpenWrt was talked about a lot. OpenWrt, if you don't know, is the firmware replacement for routers, right? If you don't trust your ISP's router, for example, which you shouldn't, you can flash OpenWrt onto it. It's a kind of mnemonic for DDWRT. It's like the old version of it. OpenWord uses Moosel C, but by default doesn't use Moosel OpenSSH.

It uses DropBear, which is not vulnerable to this. It's just, it's weird. It's another kind of, like, supply chain-y, log for J-E issue where, like, people now have all these embedded devices all over the place. And, like, they're using SSH, but it's, like... What SSH do we use? Do we use OpenSSH? Are we using Muscle OpenSSH? Do I now have to go and patch my router that I bought 20 years ago? Like, what is, obviously, you should be patching, but, like, it's a similar issue where you don't really know if you're vulnerable to this unless you have very intimate details of your devices, which a lot of people don't.

Yeah. So yeah, super weird man. Crazy bug for sure. I mean, yeah, the thing that stood out to me is like what I was talking asking John about right which was like seems pretty easy to exploit seems probably pretty widespread. Yeah granted. Alpine, probably harder to hit generally on internet-facing things than other distros, right? You know, it's usually... Alpine's like a very popular Docker base image, right?

It's like where that really starts to rack up. Yep. So I wonder if there's like... you know, just networking complexities of, like, containerized networks versus just, like, open SSH type stuff. Yeah. I'm almost curious if, like, the Alpine Docker image is not running SSH by default, if that's, like, a feature you have to enable. You know what I mean? Like, if you're running, like, a K8 cluster and, like, you use Alpine as, like, your base image, is SSH turned on by default? Maybe not. Yeah. Yeah. So the problem with, like, I think the reporting on this, too, in terms of, like, oh, we're seeing threat actors do it.

It's a really clean bug. And there isn't a lot of, like, telemetry to show you that it's happening. It's literally a password exchange over SSH. It's encrypted. So that's the thing. So that's one thing. Another thing is, like, the logs that they said the vulnerability creates or the exploit creates is literally just accepted password for account from IP from port. That's not noise. That is noise. That's not a good indicator. That's just like what's going to happen if you do password auth.

You would have to know that that account doesn't have a password, which I guess is an indicator. But you have to be very knowledgeable about how the exploit works to actually find out you're being attacked by it. Yeah. Yeah, it's weird, man. I don't know. I think what I'm saying here is that there probably is a ton of exploitation going on. It's just very difficult to detect from like the traditional ways that we do it right now. I don't like that at all. I don't like that at all. Yeah. All right. This one's weird. It feels like this is an old bug at this point because of the gap between our two recordings and how fast things move.

But I almost like being able to talk about some of these things not like day of. Because you do get to have this conversation about exploitation and things like that and complexities of environments and whatnot. So super interesting one. All right. Going... Man, I can't go... I just got back from a conference. I was speaking at a conference yesterday in Texas that I drove out to. And even the parts of the conference that weren't AI related were AI related. You know what I mean? This is all anyone's talking about.

It's the whole industry right now. Right, right, right. And so we're going to put a dollar in that jar right now. OpenAI is officially a cybersecurity vendor. There's no other way to look at it, based on a few announcements that they had recently. So what started, you know, getting onto our show notes, and I think you made a video on your channel about this this week, was that they are doing a lot of their own vulnerability research, right? So they're eating their own dog food. They have a cyber team. We've got friends on that team. Like, you know, they're doing awesome work.

They're obviously sucking up really good talent. And that good talent turns out with... you know, unlimited tokens and probably unguard railed frontier models that none of us have access to that they don't really talk about publicly is my guess. Right? I don't know. It's like Matt's speculation time. Maybe. You know? Yeah. But I think it's probably a pretty good guess, right? That they probably have versions of their own models with less specifically cyber guardrails on them than that they're even releasing to their cyber products or anything like that, right?

Yeah. But they came out with basically a thousand LPEs this week. Can you, like, A, you made a video about this. Can you, like, walk me through this? This is part of, like, their whole Patch the Planet initiative, right? that like, you know, we've been trying to talk about some of the people that run this on the podcast for a while. It turns out they're busy finding thousands of vulnerabilities, whatever. Employed. John Hammond showed up. No. So... Montana. Yeah, come on, guys. But what the hell is going on?

Can you help me understand the scope of this number of LPEs in this singular initiative as well, if you have any understanding of that? When you say scope, define that, I guess. Compared to normal. Like, compared to, like, how many LPEs do we find in a year normally, right? I have an idea. I kind of want your vibe on it. Yeah, I guess. So to frame that question, the information that I don't have is how universal are these LPEs? Okay, yeah, that's a really good point. The main question that you're probably asking is, like, is this similar to, like, Dirty Frag, or is this similar to Fragnesia in that it is a universal LPE or Dirty Cow, right?

That I don't know. But I want to highlight to the audience here, when I say 1,000 LPEs, this is not 1,000, you know, null pointer derefs in the kernel that can be taken advantage of the DELS system. I'm talking literally they had... I'm referencing a graph here from OpenAI. We can have our editor put it on the screen. But they had 2,300 candidate vulnerabilities, so suspicious blocks of code. Of those 2,300 candidates, 1,700 were actual bugs, like vulnerabilities in the kernel, of which 1,200 were LPE candidates.

948 of those turned into working LPEs, meaning you are Ubuntu on a box. I'm saying Ubuntu, but some Linux account. You are Ubuntu on a box, you run a script, you are root. That is it. And then 291 of them are not LPE pocs per se, but they're leak pocs. So they're, you know, leak privileged memory out of non-privileged memory or leak privileged memory out of non-privileged user mode execution. This is truly unprecedented. A, in the fact that it's volume, right?

Just of the number of LPEs that have fallen out of the sky at once, And two, the fact that it was automated. Okay, the fact that we have now, we're not running SAS tooling, we're not running fuzzers, we are using LLMs to drive the ship and find these vulnerabilities. And I want to ask you a question, Matt, do you think $1.2 million is a lot of money? For this many vulns? No, just in general. No. Oh, in general, I would take it. What I'm saying is that is a lot of money. That is a fuck ton of money. I want to highlight that. People are going to say, oh, it costs them 1.2 mil.

Totally agree. That's a lot of money. I don't have that to spend on fucking LPE finding for fun. No, thank you. I have a huge question here. If you divide that by a thousand, a thousand dollars per LPE is not a lot of money. No, totally. I also have a major question here. When we're reading about Frontier Labs talking about token costs. Sure, great. Like in my mind, is that an actual incurred cost for them? Because they would just be paying themselves, right? That's like their like advertised cost for the tokens used.

So for them, I don't really know the answer to this, right? Because I don't have the inside baseball, but is that like an opportunity cost? Like it, because they were taking up that like, you know, compute basically, right. They were taking up that token usage. They couldn't go be charging for that at that exact moment in time. It's not like there's a queue going on here, but you know what I mean? Like, I understand. Was there some sort of opportunity cost that, that they use their own thing that they couldn't be selling? Right. Then. Okay. That's a real number. Right.

Otherwise, I think we're talking electricity. As far as I understand it, and again, this is me guessing off of their verbiage. I am purely tea leaf readings. I don't know. Yeah. But the graph that I'm citing, they specifically said total API credits. But that's what they charge. That's what I'm saying. I think what they're insinuating there is this is the cost of the public facing API if you were to use our credits to do the same thing tomorrow. So did it cost them anything?

No. Or electricity. But this is what it would cost. It would have cost somebody else. Totally. But I just want that to be crystal clear for people because some people are like well, OpenAI paid $1.2 million to better the planet. I'm like Right. I don't think they did. No, you know, they had like a, you know, a rounding error of electricity costs to like run this initiative. But that's what it would have caused you or me to do it on the API, right? And this issue that I have with like, this is a problem that's always been a thing in volant research in general, right?

When I say the word resources here in a second, I'm going to say the word resources. That is resources of time. That is resources of money, electricity. When you're doing volant research, the inherent problem is that you know there are bugs. You just don't know where they are. So the question is, where do you spend your resources? Where do you spend your time, your electricity, your tokens? If you work at the token factory and get the tokens for free, and I know that that's not maybe what's happening, but like free generally. You know, it's a lot easier to justify let's ingest the entire Linux kernel source code, which is 40 million lines of code to find 1100 LPEs, right?

So great work. Like, obviously, it's good shit. But I agree with you that it's not like they said let's spend $1.2 million. They said let's burn some electricity on this and say we spent $1.2 million. I also wonder, this is me postulating, because if I was in their position, right? If I had a shovel at the token factory and I was standing in front of the vuln furnace, right? And I was like, let's go, right? So I've tried some vuln research. I know we were just talking about obliterated models and I haven't gone that far, but I have done some vuln research with models to be able to speak intelligently about it enough, right?

And just see what they were capable of and whatnot. I was doing some stuff with browsers I accidentally filled my hard drive. Yeah, it turns out when you're fuzzing browsers, you can make a whole lot of noise, right? The corpus gets very big. Yes, the corpus gets very big, right? So I wonder, right, like you and me kicking off like, hey, it's time to find some vulnerabilities in the Linux kernel would take up a ton of context in tokens just firing that like project off the ground, right?

Yeah. I wonder if because they've got this long-standing Pat's the Planet initiative, right, and whatever, I wonder if they have to, like, do that instantiation every time or if they have some sort of harness, for lack of a better word, right, that's like... No, we paid that once, right? Like, we have it. We have the Linux kernel in something that resembles context to us, right? Because I don't know. I don't know how differently they're interacting with the things at the token factory.

You know what I mean? Yeah, yeah, yeah. So I will highlight that in the article, they do a very good job on being a little transparent on what they're using. So they do have a special harness that they wrote that does a pretty cool thing, right? So as I highlighted... Linux kernel is 40 million lines of code. Very hard to reason about with an LLM. They had this system set up where they use these things that they call the survey agents where, hey, you're not going to go find me vulnerabilities. You're not going to go into deep thought on this. You're going to go look at the code. We're going to break it up in the subsystems and give a subsystem to every agent or group of them.

And you're just going to find me suspicious blocks. uh structures that are hard to get right or code patterns that are dangerous and just say hey this is a suspicious thing and that is what the candidate findings are i believe then from there you give that suspicious candidate to a research agent i think they call that an analyst agent to do the actual you know go deep on this put it in a vm write a pock crash the kernel like write an exploit type thing So there is some harnessing going on that I think is accessible to the average person. But I think to your point, is there like a middle layer, like before the inference but after the LLM harness itself that says like, hey, we have completely embedded.

We've contextualized and compressed the whole Linux kernel, and you're going to infer over that, not the plain text. That I don't know. Yeah, I don't know either, right? And maybe not because they are being so transparent with all of this, right? But so, I mean, it sounds a whole lot to me like when Mythos came out and we were starting to read about Project Glasswing, right? And I know we talked a lot about that at the time, that like Mozilla and Cloudflare and some of these early Glasswing participants who were able to be more transparent than maybe some other like large corporations that don't have as friendly of a PR department.

And this is what like Mozilla, this is exactly what Mozilla was doing, right? To like, you know, they were literally file by file, like giving Mythos a file saying there is blank vulnerability in this file, gaslighting it. Gaslighting it. Yeah, you have to. And just saying there is this vulnerability in this file. find it, right? And then next file, next file, next file, next vulnerability class, next vulnerability class, next vulnerability class.

So like you can imagine that like workload getting quite large, right? If you're doing a file on file. So it sounds very similar. Yeah. I was writing a harness to try to reproduce that process, and I did it to try to find one of the most recent 7-zip XZ cursor out of alignment issues. I lit $300 in token costs on fire so fast. Like, one run of the tool where I took every bug class, wrote a prompt for it, and then pointed every bug class at every file in 7-zip, which is not a lot of files, by the way, was like $100 per run.

And I was like, all right, that's enough. That's enough of that. I'm not Mozilla. I can't afford this. So yeah. We're running on that podcast money. True. Which by the way, next year is coming up guys. Like and subscribe. Like and subscribe. The other dovetail off of this, right? That like, okay, OpenAI obviously doing cyber stuff. So like the Pats of Planet initiative got it. They're like trying to do all this stuff in Linux and open source, you know, project, major open source projects, stuff like this. But they did announce this week

that they are just straight up competing down the line with uh cyber security vendors right they they launched daybreak basically red and blue right yeah uh daybreak defense daybreak uh red team And you can now hit them up and be like, let's do this. Like, I would like to hire you to be my offensive security agent. Right. Yeah. I find this super interesting because I think we all saw this coming. Like, there was just no way that they weren't going to turn on the money stream of cyber stuff when like you go to the black hat vendor floor and it's like, you know, a gazillion dollars.

being thrown around saying that they do this kind of stuff and they're all just front-ending the frontier model capabilities anyway. So it makes sense that they were going to do this. This is the number one question I ask. We've done a few interviews now with AI security vendors. This is my number one question now. It's like, how do you continue to do what you do when the frontier models are going to start doing more and more of this out of the box? Not even this cyber offering that I'm talking about, right? Like... you know, Opus can do some of this stuff for 20 bucks, right?

Like that vendors were doing as part of like their enterprise deals just a few years ago, right? So it's like, how do you kind of stay ahead of this as a cybersecurity industry, right? Yeah. It's difficult. Geohot, I think, said this in an interview talking about, this is like back in like the Sonnet 3, 7 days. But he was like, yeah, there's a problem where you have, he calls every company a tier X as a function of like what thing they provide to the market. So like tier zero is like the data center, right? Like they are the hardware, they are fundamental, and they are very hard to replace because data centers are expensive.

And then you have like the tier ones that are like, ugh. open AI or Anthropic. They provide the frontier models and you kind of have like the tier twos and threes that all kind of just wrap each other. And his argument was that like the tier ones are going to just consume all of the other companies because they own the thing. You know what I mean? Like the data centers aren't going to become frontier labs, but the frontier labs can kind of become whatever company they want because they have the models that can do pretty much anything So it does beg this kind of scary question about like, how far do these companies go if they can just have a model that runs a company that, you know, does it better than the average competitor?

It's weird. I'm not sure where it goes. I mean, this is not a cyber-only conversation. No, not at all. Like, we just saw Claude Meetings come out yesterday. And it's like the laundry list of startups that do like AI meeting notes type things. And it's just like, oh, well, now it's just in Claude. Like, that's just there now. Right. It's going to get weird. It's going to get weird here for a bit. But I thought it was just super fascinating. We don't have a ton of details about like their cyber offering out of OpenAI. I'm interested.

Obviously, the red team stuff is basically everything we're talking about. Right. Like that we've been talking about is like, do you want to find bugs? Like, cool. We got a bug finding super, you know, machine here. Right. The blue team stuff. It's not like they are going into the AI sock space or SIM space or anything like that. It's just fix the bugs that we just found. Right. So they're just staying close to the code right now. So I think that that's a bit of runway, right? If you're just an AI bug finder, you're probably not happy about this, right? Bug finder, bug fixer.

But I think like there's a laundry list of companies that are doing more security operations, you know, either offensively or defensively, right? But just like more into the security team with AI. And I don't think that this announcement really impacts that much at all, right? Yeah. It's not like you're hiring them to, you're not going to go to API, OpenAI directly to, you know, run your SOC alerts, right? Off of daybreak type stuff. It seems like mostly, hey, they have this model.

And like, I want to be clear until very recently, a lot of these models were not even intentionally trained with any additional cyber intelligence. It's just they got continually better at code and engineering. that, you know, so much so that like stuff like memory bugs and injection bugs and things like that that are very, very syntactical and like, you know, whatever, that that's okay, now we're just playing in their playground at this point. weird stuff super weird um speaking of injection bugs and uh code that just should you know someone should just go over to fortinet and be like we're gonna one shot rewrite everything like let's do it like rewrite fortinet in rust make no mistakes how do you think do you think we would be better off or or what do you i don't know bro i actually highlighted that in my um open ai thousand lpe's video like

Does Rust even matter anymore? Because if the AIs are so good at coding, do the promises that Rust offers matter? You know what I mean? Versus just like the mem copy where N is a network value can't happen. Are you saying? Are you saying? Hold on. Are you saying that we invented entire programming languages that were memory safe because of skill issues with other programming languages? Yes. I am. I am saying that. And people are like, oh, it's only a skill issue. Like, just be good. I'm like, okay, cool.

Microsoft would love your opinion on that. That's awesome. I'm sure that they would love to hear just hire better programmers. Good fucking idea, dude. Yeah, it's been 30-something years of C code. Yeah, surely we'll figure it out one of these decades. just be better at sea oh okay so we're only going to have the best that makes perfect sense this Fortinet to shell vulnerability is interesting because when I saw Fortinet I was like oh god it's another endpoint device special on the lowdown this week interestingly enough there's actually a vulnerability in FortiClient which as far as I understand it is the client that you use to connect to an Scell VPN so Fortinet

They're a device manufacturer where you can use their endpoint to act as like a VPN concentrator, right? And you can VPN into like your corporate network and do HR work from your house. But you have to go inside the network to do it. The vulnerability actually allows a malicious website to pop calc or whatever code execution on your computer because of vulnerability in the FortiPAM portable authentication module inside of FortiClient. Right. So there's literally a POC, not really a POC, but like a video of the, the, the person who found this executing the vulnerability where literally they just have for the client with for the PEM on their, their device, on their little home computer.

They go to a website, pop calc, GG. That's insane. Now, what I don't know is if this is a command injection vulnerability, where just, like, you do shell execute a calc.exe, or there's actually, like, some ROP, some, you know, stuff going on, some memory corruption that actually they have to kind of tailor this to the exact target. But regardless of that, it's a really kind of neat vulnerability where, you know, the era of browser go to website, get a shell on your computer is alive and well, despite memory safety in V8, you know?

Man, pretty weird. I like I reiterate my my point earlier. Just take the two shell suffix out back. And I literally brought this up and I was like, guess what they called it? Guess what they called it, low level. Yeah, yeah, yeah. I literally as a joke, I was like, oh, is it Fortinet to shell? He's like, yeah. So the other part, the other part that I thought was kind of funny about this is, did you say that they're just getting out of the SSL VPN business? Yeah, so I Googled, just to get some background context on this, I Googled Fortinet SSL VPN, and they said, yeah, as of May of 2026, we are no longer going to provide SSL VPN services.

So they just are not going to run that part of the company anymore. Moon's haunted. They were like, you know what? We tried. We tried. We tried to go back, baby. Good effort. Turns out really hard to continue to do this without pumping zero days out. You know, okay. I want to pause here for a second because we dunk, we dunk, we dunk, we dunk, we dunk, we dunk. And by the way, more dunks are coming. More dunks are coming. That's fine. Yeah. I have like a serious, I want your serious opinion about this. Okay. For these types of companies, right?

And I know that neither of us are like qualified to actually answer the question I'm about to pose to you. So like, okay, whatever. Right. But like, What is stopping them? The vulnerabilities coming out of these edge device companies, right? Like the Cisco SQL injection one in the email gateway that we talked about a few weeks ago. Fortinet constantly, right? What's the other one? Avanti constantly, right? Like all of these like edge devices. What do you think is stopping them from seriously? Like the types of vulnerabilities that we are finding are things that literally AI can like just will not code anymore, right?

Like you said. What are we waiting for? If you're a publicly traded company with massive deployments of edge devices, what are we waiting for? Before I answer that, I will say I've been saying a lot that like AI can't produce these code patterns. I actually pretty recently had AI code a really trivial path traversal bug like for me on an actual product that I'm making. So it is possible. I'm not actually sure how it managed to pull this off, but it did do that. So we'll just kind of it's less likely that an AI will do it than a human, but it is still possible.

Now, your question is, like, why are they not just auditing these products or rewriting them? I mean, man, you've ran security teams. I've ran security teams. I think it's literally just a manning issue. Like, they don't have enough people to be able to give a shit about all their products. Because at the end of the day, like, you only have 10 people. And if you want 11 people, you've got to get more money. And if you need more money, you've got to convince the CFO and the CEO to give you more money. I just don't think these companies are managed in a way where there are enough people that can do the thing they need to do. You're right. You're right. And I know this.

And that's why I said, we're not even qualified because it's not that they don't know how to. It's not that they don't have the tools to do it, right? It is financial, political, business priorities, all this kind of stuff. But if you're me, right? And I'm the head of a publicly traded company that has these edge devices that are supposed to do security for people out there. I like... I have a really hard time trying to figure out what else my priority would be.

I could see a ship zero days in my security edge devices. I could see a world where you enter kind of cognitive dissonance and you're like, ah, the AI super hacker, nothing we can do. You know what I mean? Like maybe, man, maybe it's just like, I'm not saying for anybody is like this. I'm just saying when you have so many competing priorities and you have to make X people, happy shareholders, board, whatever, um, you know, you probably find ways to justify the fact that you're not doing all of your priorities. Yeah, it's money. It's money. I'm just being cynical.

No, I'm with you. I mean, it sucks. We all know this, right? We've highlighted this a lot. The technical solution is not the solution because it's always more complicated than that, unfortunately. And not to say that, like... zero days are going to stop existing. Like, you know, not to say that we're going to get to the bottom of this. Right. But it's just like the vulns that we're talking about, man. Yeah. You know what I mean? It's just stuff that like pre-AI shouldn't have happened either. Right. Like, it's not like, oh, you know, this is H1. That's, that's just like kind of funny and crazy.

And like, it's been, you know, it, that whole story took decades to unfold. So like stub our toe on that bone. Right. It's not like, oh my God, how did we do this? That's not, you know what I mean? You can see very clearly how that happened. Totally, totally, totally. Even though it's a major bug, I'm not like dunking on it. Like that's just crazy, right? But like, man, some of these, right, that like we keep talking about in these literally, you know, multi-billion dollar publicly traded cybersecurity companies. Yeah, but what was the one that was like an SQL injection from a field in an email in an email security gateway?

Yeah. Like how does that happen? Yeah. Right. Or like how many of these public management interface login portals that like pre-off? Like, how do you not at this point, I don't care, like budget staffing, all this kind of stuff. How do you not have a frontier model just banging on your public facing management portal stuff? Right. Sure. Most of your code is beyond that. Like, most of the code of these products is beyond that public-facing thing, right? Like, you know, most of the functionality that could be reached by an attacker is beyond that auth.

And it's like, okay, whatever, right? Like, we're going to find vulns in there forever. But, like, how are you not just, like, throwing something at this publicly-facing part of all of these products right now? Yeah, so some platforms are hard to fuzz, right? Like, embedded stuff in particular is hard to fuzz because of, like, the... hardware peripherals it may depend on. Like in the Citrix case, right, they're literally an application gateway, so their performance is super important to them. So the way they architect it is that, like, all the code goes for the kernel first, and, like, some of it's dispatched out.

Maybe that's hard to fuzz. Maybe their tool chain is super old. Like, if you look at it, it was running FreeBSD, and, like, PIE wasn't enabled, and, like, ASLR wasn't on. So to me, that just implies very old tool chain, A company that sells a product is not incentivized to update their tool chain, which means that, A, the security mitigations aren't enabled, and, B, it's very hard to get things like ASAN or LibFuster to run on those tool chains. So it's just like traditional corporate inertia, traditional corporate bureaucracy kind of set the stage for those major issues, right? Yeah, I know I'm going to get dunked on for this section, but I'm just, it's hard, man, when we do this for a living, right?

And we're just like looking at all these bugs going like, what's this one? This bug? Like, are we, you know, what's going on? All right, you want to dunk one more time? Yeah, dude. The number of the year? Four, baby. New meme number just dropped. Okay, so. A mud rub, right? This happened when we were at OneCon, Sentinel One's conference. Yeah. Basically, it turned out that the private keys, plural, for eight intermediate certs for a certificate authority, eMudra, were single-digit integers, one of them being four.

Yeah. Fucking hilarious. The question becomes, how does this happen? How is it possible that the cert number that the private key is for? A little bit of detail came out, so we don't have the exact script that went down, right? But basically what happened is, and the reason why I'm highlighting this one page, I'm going to have the editor put it on the screen. It's like Emudra's FAQ for their HSM security. How does Emudra integrate FIPS-certified HSMs to protect private keys and ensure crypto agility for post-quantum cryptography algorithm rollouts?

If you know anything about FIPS, which I really didn't until I read more into this. So FIPS is the Federal Information Processing Standard, which basically means if a device is FIPS compliant, it means it is allowed to process government data. It is also internationally recognized as like if you are FIPS compliant, you can be a CA, for example. Right. Being FIPS compliant means that you implement the NIST guidelines, the National Institute for Standards and something. Right. On cryptography, okay? One of the FIPS requirements is that you're not able to export in plain text the private key for an HSM, okay?

And the HSM that Imudra was using had FIPS mode enabled. Isn't that great, Matt? Guys, look, they're FIPS compliant. How could this possibly have happened? Matt, they disabled FIPS mode And exported the private key in plain text, which is fair because the script they were going to use, Matt, they were going to encrypt the private key for ingestion into a new hardware security module. The vulnerability that happened, the issue that happened was something in their script. We don't have the details. We don't have the script. Something in the script truncated the key.

So it read out the whole key, all N bytes, right? And it made it so the key was only the last byte of the key. So in the case we talked about, the last byte of the key was four. That's the new private key. And they ingested that into their new HSM. And then the new HSM was fifth mode. point of an HSM is that the key doesn't leave the HSM. Correct. That's the whole point. I'm shocked the HSM doesn't have a way to export it encrypted.

How is that not a thing? It sounds like they got around that. It sounds like they disabled that, right? So why would you, though? Like, again, I have never, like, physically put my hands on an HSM. I can't speak to this. But as far as I understand it, it's meant to ingest and cryptographically secure the key material so that no one can see it. And that if it ever comes out, it comes out encrypted and signed. So I would imagine HSM has that functionality, right? And that's why the NIST guidelines are literally, hey, you can export private keys, but they have to come out encrypted.

There's no way that HSMs don't do that. Maybe it's because they're going between HSM vendors, which is part of the write-up by Mozilla. Like, it's going from vendor A to vendor B, and, like, maybe the export file format isn't correct. So, like, they have to do it themselves. I don't know. Yeah, and obviously they were thinking that they were going to do something with some Python script to manipulate it, and they couldn't manipulate it if it was encrypted. Right, right, right. But what is the thought process here? I just don't even understand the thought process. Why even spend all the money on vendor A or vendor B HSMs if you're going to do this?

This defeats the purpose of the HSMs. It is funny, though, but, like, I was about to say, why didn't they look at the private key and check? Well, actually, they did the right thing by not, like, printing the private key, right? Yeah. They, like... But they just trusted it. Yeah, right. I mean, maybe you print land of private key. I don't know, but... At least look at it when you're devving that Python tool that you were doing, right? And be like, all right, did I accidentally truncate this to one character? That is the frustrating part. It implies that there wasn't robust testing of that Python script.

At all. Yeah. Not robust. They didn't see that the output of a single character is crazy, right? Never mind look at it. How did the unit test not say, hey, by the way, the output of this is not an effing key? Yeah. To our point about AI before, too, there is an argument to be made about, hey, maybe we don't write scripts that deal with CAs, with AI for privacy concerns of the key material getting sucked into the model. But at the same time, an AI would have caught this. Well, you can write the script. It's just not then... You don't want to be using AI in line for actual key material stuff.

But you can use a model to write this script Why does the script exist? What are we doing? I don't even understand what they're doing. The whole point of HSM is that that's it. That no one can see the key material because it's still in there. To turn off FIPS mode and then do that is just insane. Further dunk. I'm not trying to dunk per se. The whole situation is so unbelievable that everything that I'm going to say comes off as a dunk, so my bad.

Regardless of our dunking, Emudra is no longer in the certificate authority. It's a second in a row. Two vendors in a row that have failed so miserably. They're like, you know what? We're out. We don't care. We're done. You know what? This whole TLS certificate authority thing for us, not for us. We're good. Check it out. Moon was haunted. We're coming home. Moon's haunted. We don't even want to go back. That's not a thing. Moon's haunted. So interesting week for sure. I think we mentioned this with John, but it's just like, it's funny how only two 10.0s came out this week, so it's kind of a slow week for us.

You know, like, all right. The most important part of the podcast, Ed. Oh, no. We already did the vendor cut, or the sponsor cut. Nope. Most important part. Sorry, yes. Second most important part of the podcast. Okay. Alliance or Horde? Oh. On PvP, Horde, buddy. Okay. Okay. I'm like spiritually Horde, but I got bullied into playing an Alliance character on forever. Ironically, I'm the opposite. My fondest gaming memories are like being at Sentinel Hill when I was like 14.

Yeah. So I missed WoW in my teenage years. I played a little bit during COVID. Some friends were like, we should raid. What was the expansion that came out? Battle for Azeroth, I think it was called. So I played a little Battle for Azeroth, which I know is way late. So I missed all the Wrath of the Lich King stuff, Burning Crusade, all the stuff that you guys all look back on fondly, I completely missed out on. But... I do have enough homies playing forever that I was like, all right, I'll download forever. So I've been putzing around with it. But hilariously, all of my algorithms have me pinned because all the videos that I'm getting recommended are wow for busy dads.

And like wow for losers who can't actually play wow. You know, and I was like, oh God, how did they know? How did they? I didn't, I didn't search that. Like it is true, dude. Like I thought I would be hungry for like the classic feel of like the OG game. And then like, I've got, I've got kiddos and this job. And I realized like, Oh my God, when I was a kid, I would just degen and play for like 10 hours a day for weeks. Yeah. That's all we did. You can't do it. You can't do it. I can't do that anymore. I will say, I don't know. You know, I didn't really have that experience.

Um, I did do that a bit in Final Fantasy XIV again during COVID, right? I just ruined my sleep schedule and got really into that as an MMO. It's the only MMO that I really got like super, super, super into. But I will say, I don't know, Forever feels like there's not a lot of that pressure. It feels a little cozier. It feels like everyone's just trying to go around and kick around. I see a lot of people like fishing, you know what I mean? Like... Yeah. So the problem with retail and with WoW in general is that as the internet got more mature in like the early 2010s, the game got boiled down into a science where like if you were not doing the meta thing, you were not leveling the fastest, you didn't have the best gear, you were an idiot and like shouldn't play the game.

And I think WoW Forever is an attempt to return to the fact that it's like a game. You know what I mean? It's not a job. It's not that serious. Like, if it takes you four months to hit level 40 or whatever, like, that's okay. Because the game's going to be here for a long time. The idea is they're going away from vertical progression where, like, you've got to build, build, build, build, build, to, like, okay, cool. Like, this expansion comes out now. Maybe another one comes out later. But, like, they don't depend on each other. You know what I mean? Yeah. Which I think makes it closer. Yeah. Yeah, there was this pressure to get best in slot.

Sorry if you guys are not getting... I got so many DMs about... I made a slight joke about... Because I went and spoke at that conference yesterday. Yeah. And I said, I logged off for WoW for this. I got so many DMs about it. I was like, oh my God. Our audience absolutely is on WoW forever for sure, right? Yeah. But yeah, like I'm like down for the cozy experience too, right? I've got some add-ons from Curse Forge like to help guide me around and level. But like I'm down to just like take little side quests too. I'm not like trying to get the most efficient thing possible.

And then like I don't even know what best in slot stuff exists for anything right now. You know, I have no idea. So like and that's kind of cool. I'm just trying to like this, like at this stage in life, I'm trying, I don't want to have a ton of gaming time. So when I do, it's like, can I be on discord with some friends that live all over the world? And like, can we play a game together for a little while? Then that's cool. That's valuable. Right. That's like actual useful time, not D gen time. Right. That's like, Oh, I'm connecting with my homies that live nowhere near me. And this is great. Yeah.

As a man in his 30s, the most fun that I've had so far has not been, like, getting a BIS item in retail WoW. It's been, like, getting blown up with my friends in Outlaws. You know what I mean? Like, that's the fun shit. Yeah. That's a good time. I like a grind. I'm, like, if I like gaming, I like a grind game, right? So I like that I'm working towards something. You know, it's like, yeah, okay, let me, like, run this over and over again for a year. And then, like, you know, whatever. I'm kind of good with that. So, yeah. Anyway, I got, I got enough inbound, but I was like, we got to talk about, wow.

Cause I know you're playing a little bit too. We're both, we're both filthy casuals. I think you're 13 on like 17 or something. Yeah. So, uh, you know, I, I've been traveling a bunch, you know, I've, I've logged on maybe two or three nights with some, with some friends and ran some quests. So it hasn't ruined my, I haven't let it ruin my sleep schedule yet. Yeah. Maybe when we start raiding, I'm going to, I'm of course going to roll something that we can play together. Or you're going to roll something so we can play together too. But yeah, maybe we can start raiding or something. Because your ally PvP, you said?

PVE. Oh, your poor PvP. But we can play together. Yeah. PvE Alliance. So yeah. Okay. Cool. Sweet. All right. That's it. Take us home, man. As usual, we appreciate it. Guys, we're wrapping up 2027 here in a little bit. You know, it's been a good couple months with the podcast in our first year. Episode 20. And we really couldn't – oh, yeah. This is number 20. This is the 20th episode. Guys, we really – a serious talk for a second. We could not have done this podcast, started this whole thing out without the support from you guys, our listeners. Really do appreciate it.

Next year – It's going to be a stellar year for us. We're going to have a lot of cool guests on here, doing a lot of cool events, maybe even hosting an event ourself. That's kind of in the cards right now. So if you want to be in the loop on all that stuff, hit that sub button if you're on YouTube. And if you're on Spotify to actually get recommended us again, rate us five stars. It actually helps push us out not only to you, but other listeners that are potentially interested as well. And maybe you're a sponsor of the podcast. We do appreciate your support on getting the team together and making this bad boy happen. So appreciate y'all. And Apple too.

You know, I think we're starting to see the snowball pick up on those audio platforms. YouTube's weird and spiky and algorithmic and like the other ones, it like really takes you guys helping and supporting and whatever. Cause there's no discoverability there. So really appreciate it. Yeah. Yeah. Yeah. All right. Cool. Happy hacking. Peace.

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. 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

  3. 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

  4. 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

  5. Cybersecurity Experts BREAK DOWN the shocking OpenAI Spying Incident, Crypto Backdoors, AND MORE

    Oct 19, 20261 hr 17 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