Security Is Getting SCARY

All right, before we begin the stand-up, we do have a bit of a blocker. I've noticed that some of us are wearing costumes. Trash, love the costume. Hey, you know, this is a great costume for me. Yeah, party on. TJ said he's going to reveal his, and I must say, Casey, where's your costume? Very disappointing.

Very disappointing. Yeah, you know, when I was working on that graphics project that I work on that I can't think of the name Vulcan, when I was working on Vulcan, I got too busy. I'm going to wear this, but I'm going to be a skeleton for Halloween as I peruse about the efficiencies of game code. You know, that's what I'm working on right now.

So it'll come like probably today or tomorrow. Okay, that's good to know, because I'd be a little disappointed in you, Casey, if this was the best that you had. Yeah, no, it's upstairs. I'm working on it. Are you ready for my costume, Brian? I'm ready. Okay. Today, it's a very well-known costume now, and so I am dressing up as what might be the most cryptographically secure number of all time.

No way. Let's go. I am the root certificate. You can trust me. You can trust me. You can trust me right here. That's all you need to know. But also fantastic forward is that too. That's a great four. That's a good looking four. It is a good looking four. It's like parallel lines, even perpendicular. Oh, chef's kissed.

I have to give a shout out to my wife who free handed this about 12 minutes ago. Wait, wait, wait. What do you mean free handed? She just drew it on here. No, it's a marker. It's a marker. That's a marker? There's no way. I thought that was tape. Dude, that is incredible. Okay, that's pretty good. I'm pretty solid for you having.

No, I'm impressed. Ed, I'll see if I can find it right now. We may have to bring it back momentarily, like later in the episode. My wife did at one point, it was a little bit of tracing as well, but she did draw a QR code for me. No way, dude. Yeah. That's to your telling. Some tracing involved. Oh, it took her a little while.

I tried, I think, three times, and it failed every single time. No matter how much I tried to trace it, it didn't work. I feel like I could one-shot a QR code. It's a good challenge. Take a video after this Prime and let's see it. I'm ready. He's going to cut it like a dream. I'll find it. I'll find it though.

So she's got some skills. She's got some skills. And I'm not even afraid to say it. She's pretty cool. I like her. I hope so. I'm glad you're not afraid to say that DJ. Well, you know. Dangerous words in today's day and age. Liking your wife. How could you honestly? How could you? I'm going to find the QR code because I want her to get the credit for it.

I'll find it someday. It was on a sticky note too, right? I found it. I found it right here. What was the QR code? Oh, Ed, great question. Let me... Let's pull my phone out. I'm getting ready. I know. I'm getting ready myself. Prime, you show it. You show it, Prime. You show it on your screen here. I'll send it in studio chat here so you can show it to people.

Don't send it in studio. Send it in Discord. Okay, I'll send it in Discord. Wow, she drew this? Alright, hold on. Oh, I love that we're edging people. People are watching at this exact moment and we haven't cut this segment out. Wow. I've got 16 updates for Discord, guys. I've got 16 Discord updates downloading and installing right now.

So probably unrelated to what the topic of today's episode is. That's probably not why there's 16 Discord updates to download. Yes. No, completely unrelated for sure. Say you're alive. It's alive. It's alive. It's alive. Franky Prime? Yes, T-Gore? Why did Seer by Sentry? To submit a PR for a chainsaw service.

I've been adding a little new skill to Prime Claw. Wait, who made the PR? Like I said, Seer by Sentry submitted it automatically. It can not only catch the bugs, it also fixes them for us. Turn around, turn around. I should add helicopter blades. Baby, you're gonna see the world. What are you going? I don't want her to get hurt.

First, it's Tijko. For the last time, Seer by Sentry can keep her safe. Even the most dangerous code, Seer by Sentry, and submit bug fixes without us even knowing that there's an error. It's almost like it's self-healing. Oh, I've got a brand new idea. Self-healing. Don't wait until the project turns into a monster.

Automatically find and fix issues the moment they appear with Seer. it's should have been on the website i should have been on good oh i probably that would have been easy i i'm booting up a native app when i was already in the browser by native electron it's literally just here you you you bring the browser good point good point uh there you go prime you can show it now all right Look at that, dude.

So this was also right afterwards. Oh, yes. Yes. Claps. Claps for Kayleen's QR code. But so there was a post where a guy tried to ask out a girl on a sticky note. Impressive. And DMCA'd. All right. That's fine. That's one where I'm always willing to give him the money. You know what I'm saying? Yeah.

Yeah, I'm willing to give him the money. I'm willing to take the DMCA on that one. You say, this ad revenue, it is yours. We might not even be in the game if it weren't for you. You know? True. Think about it. How many videos does that video have? Is that like an $8 million video for him? I don't know.

This one has 1.8 billion views. So I don't know what that equals, but that's a lot of billions. That's a lot of, that's a lot of views. That's insane. It's something. Almost two billions. For sure. Thanks trash. Almost two billions. Just in case. Okay. So anyway, sorry. Yes. So this was free handed with a Sharpie marker just before this meeting.

So, so good. Impressive. Impressive. Thank you. Thanks. Very impressive. All right. Happy Halloween, everybody. Happy Halloween. All right. Don't forget to look for vulnerabilities in your kid's candy, okay? You never know what's going to be inside of them these days. That's true. All right. So it's kind of funny that this is a Halloween episode because it kind of fits, you know, with the theme a little bit.

So, Ed, we brought you on because I think a lot of people are generally asking. The title. What the heck is going on here? Because it is a little outrageous. Let me just give you a quick recap of the last, let's just say, few days. October 20th, PHP to Shell, PHP vulnerability exploded the internet. October 20th, 4, which TJ is currently dressed up as.

That happened. October 20th, Java vulnerability with Jackson, data by an emergency Java update. October 20th, HTTP 3 DOS attack via QPAC. Very awesome. I think there was also a DOS attack via the other pack, HPAC inside of H2. Let's see. October 21st, Railway hack. October 21st, Node 2 shell vulnerability.

October 21st, Light LLM supply chain attack. October 21st, GitLab CVSS 10 RCE omega stuff if you're hosting your own stuff if you finally said enough github i'm gonna host my own stuff you would in fact be having level 10 all the way the greatest wow forever uh level hacking here going on october 21st kernel uh linux kernel accidentally leaked several like really high that was so funny That one.

That's tough. All they had to do is put the prize in the bag, man, at that point. Put the tokens in the claw, but instead they just threw it out the window. October 23rd, Ruby Regex came out, which apparently this is behind GitLab's RCE. October 23rd, PHP again. October 23rd, Python vulnerability with pillow image RCE.

October 23rd, OpenAI's model escape its offline sandbox, and apparently they found zero days in VM escapes. tail scale had tail worm on october 25th 26th we had u.s fece filing system shutdown which is the federal election commission they were getting hacked a week before the midterms that's crazy not great uh october we should keep on a scale of one not great definitely not great yeah This one is the most scary of them all.

Let's see. October 26th, NPM Worm, Armed 42 GitHub Repos. October 28th, Claude builds remote root exploit for Linux via a simple TCP use after free bug with a zero click remote root exploit. Yeah, man. And this was all in what? This is nine days or this is eight days. Yeah. Happy Cybersecurity Awareness Month, guys.

Are we aware? Are we feeling aware? I'm aware. I'm aware now. I'm feeling aware. Actoberfest, am I right, boys? True. This is Vault-A-Vault-Oberfest. Anyways, that didn't come out as much. Really good. Big fan of that. Party on, everybody. Just really came out there for that one. Ed, we obviously brought you on because you know a thing or two about security.

It feels like really high level vulnerabilities being discovered, feel like they've all happened really fast. Like, is this normal or, you know, because I also didn't include the 300 and some security bug updates, including six or 13 sandbox escapes from Firefox, six sandbox escapes from Chrome, right?

Like there's just so much stuff that was in this list. I actually had to say no to 80% of it. Yeah. What is going on? Why does this seem like, The craziest day. Is it obliterated models? Is it GLM actually that good? Is it really AI being that good? Or were these, is this all just like purely happenstance, oopsie daisies and everything just lined up?

No. So this a thousand percent is a result of AI. Okay. Now I want to be very clear. I think this time last year, I was an AI skeptic. I have been an AI skeptic for a very long time. And I am, I am slowly more and more becoming AI pilled as it comes to security. The reason being, the reason why this always is happening, I would say there's a difference between being an AI skeptic of saying the current models can do XYZ versus saying it is not possible for AI to ever get good at something.

I would say probably most of our positions on this, like in this podcast, at least right now, are not that AI was never, ever going to be able to solve any of these, but just that like a year ago, people were saying, It can hack literally everything. And clearly it couldn't because it didn't happen. And now it can.

Right. I would say that's a little different. Yeah. No, I agree. I think I may have actually even overtly fallen into the, it will never category was looking back. It's a stupid position to take obviously. But yeah, man, it's, it's getting good. I mean, I think we're seeing people awakening to the fact that if you get the right open source model, I think GLM 5.3 is kind of the go to right now because it's on par with Mythos, but doesn't have the same guardrails.

You can point it at kind of anything and it will find problems with the anything. Vulnerability research, security research is really just like advanced pattern matching, right? People that are in this space know this thing called code smell, where you can kind of very quickly look at a code base and be like, all right, they're not checking these return values.

They don't border these values from a network packet, right? Yeah. And because of that, you can very quickly use an AI to search for those patterns and automate the process that a skilled reverse engineer could already have done, right? So people that are really good at VR and RE are getting a lot better.

Like I said, I made a video on this. A thing that used to take me a week of reverse engineering is now a 15-minute task. And now also people that never would have been able to, or never really were interested in dedicating the time to learning assembly, learning what like C's ABIs are. You don't have to learn that stuff anymore.

You should, by the way, I'm not saying you shouldn't, but you can literally there's, there's like the viral reverse engineer, anything repo that's out right now. You can point that repo at pretty much anything pending like hardware DRM. And it will take it apart and tell you how it works. Right. So I think the proliferation of, um, AI-enabled RE is kind of making vuln discovery much easier.

A buddy of mine, Daniel Meisler, said, I think what's happening is we've been living in a dark room that's full of snakes and cockroaches, and we just invented the flashlight, right? Like, I think we're just now going to start to see all of the bad code. And now that there are bad coders, like, I think everyone writing these projects are good programmers.

It's just... the nature of compiled systems level languages other than go, like make it very easy to write memory bugs. So that's kind of the TLDR on the situation. All right. So if I'm hearing you correctly, this, sorry, there was a lot of things said there, a lot of little, little, you know, different paths we could go on, but I'm pretty sure what I just heard is go on top.

Is that what I'm hearing right now? Because I mean, that's, that's pretty fantastic. Yeah, dude, dude, it has a comparable, uh, Speeds, right? There's the garbage collector, obviously. It's easier to write than Rust, and it's nay impossible, pending like a nil you don't check, to have a vulnerability in a Go program.

Obviously, like, logic conditions, right? But like, memory corruption vulnerabilities are much harder to do. So go on top, man. I'm a rust guy personally, but I think go might be winning the race right now. People are just so averted to rust. They just don't like it. I don't know why, but you know, one of the things that's been interesting of like exploring different projects with like AI and how, so like I'm, you guys can see right here.

Actually I've got the, what is that? That's my, so that's my remarkable paper pro right now. It's got a, like a custom Zig, like runtime basically that I've been building. I don't know, like exploring, you know, I'm not doing too much with the actual code right now. We'll see if eventually it makes sense to look at it more or whatever.

But like part of the nice thing for that and doing it in Zig is like, it compiles pretty fast. It's nice that I can like do it. And just even today I was doing, I was like, yo, what? We need to be able to get like literally it matters for agents 10 milliseconds versus 50 milliseconds. If you're going to run it a thousand times in a day.

Oh, sure. Right. You know, like it actually like if you can get those really tight loops. So I do think there is something for like, you know, go compiling really fast and being able to have like a tight loop for some of these things and being like, cool. Like if the LLM gets to try a hundred, not a hundred times more in a day, but you know, but like literally like an order of magnitude more times or something, maybe right.

Approximately in a day to like, go get something that is actually like a really big win. Like it's actually really meaningful difference. So it'd be interesting to see like how much that, Maybe someone will invent, like, how they've got 30 new TypeScript compilers that are all 100 times as fast. Maybe we'll get a new Rust compiler that's also 100 times as fast, and then this problem will go over.

Bro doesn't even know. There is one. Don't people keep trying to, like, make languages that are new and are, like, AI optimized? Didn't, like, Vercel release a new language or something that was, like, for AI? I mean, so, hey, people making languages, Ed, is really smart, okay? All right? So don't worry.

I was going to say, that's dumb as hell, but. No, no, because there might be someone on this call actively working on this program. And he's intelligent. Anybody here might be working on it. So I do think people did release a language, but theirs wasn't released for. I think it was for agents to be able to optimize and minimally, like efficiently communicate programming concepts.

I don't know. I don't remember reading anything about compile times or any of the other facets being like. The North Star. Okay. It was like token efficient for like the code syntax. And they were just, I think, exploring. It was in their labs. I think they were just exploring some concepts there. I think they've leaned into a bunch more of trying to figure out a bunch of stuff around TypeScript land to just have TypeScript do stuff faster.

I mean, like Bunn is trying to add ahead of time compilation and stuff too, right? So it's like... I do think... Not just because I'm exploring making my own language partially for fun and partially to test out the tools and other stuff like that. But like it does seem like there are different tradeoffs for agents versus people.

And you would want to make different like design decisions for a language that like then you might for humans, even just some stuff like. Yeah. So OCaml has this concept that I'm stealing, which is like they're like header files, right? Except, you know, more expressive and you can put more things in them.

And like the type system is a little more powerful than C's. But you can actually like... do a bunch of stuff with those header files like ahead of time to compile to check if your language, like if the thing is going to work, right? Like do the types actually match? Not like the function may be bad or not implemented yet or whatever.

I think there's worlds where like we could explore languages where like you do a bunch of things just with the types of a program first until it gets those done. And those are instant to compile compared to the implementation, right? And then you just say, cool, LLMs can go fill in the implementations.

We already know what the contracts are, right? But for people, that would be so annoying. You know what I'm saying? No one is sitting there and writing all these weird commands to run just the compile time step of half of my language to do this and other stuff. But agents are pretty chill to just be like, Cool.

Go figure out the types for two days until that works. And then we'll spin up a thousand agents to write the functions. That's not insane. That's not. You're almost doing like test driven development, dare I say, but like type driven development, right? Type driven development. Specify the contracts between the functions and then go deal with the rest of it.

Right. And you just see like. I think Dylan does that or was doing that. Dylan has some cool stuff with like almost like a call stack thing where it like writes out what the call stack will be. And so you can look at what services are depending on other ones. And so it like writes almost like a trace that you would get if you were actually using like Otel for it.

And then like, no, no, no. Like this, like I see this thing and way down here is an HTTP call. No, no, no, no, no, no, no, no. We cannot be doing that here. That has to happen way out here and get piped into our, you know, so you can see like at a higher level. Right. Like what's, what's going on there.

So I think there are like... It makes sense that if the constraints and the things that like LLMs mess up on or like the kinds of things where you would like to have like a lot more assurance of, we should like figure out how to push those further into the compiler. It seems like to me like that makes sense.

Like that's why I like writing. Well, I'm not really writing it that much, you know, to be honest, of course, but like like effect for TypeScript is good because it just tells you like, does this thing depend on. the database, right? Or like, what kind of errors can this thing have? And are we going to handle them?

That's like, I want to know that. I don't want to have to trust that people remember what exceptions get thrown by functions seven layered steep. That might work when you have a team and you're all working in the same code base all the time, but like not when you like, resurrect a little agent from the dead every single time and it has no context on what's going on in your code base like that doesn't sure sure you know right it's like those kinds of things sure okay cool i can pollute the context a ton sweet now i can just put a hundred thousand tokens to to waste every time i start an agent that's smart that's good that's what we you know versus the language being like expressive about the exceptions it can throw

Right. Yeah. So like the language that I'm exploring, like right now, it has like tracked exceptions, basically, except there are errors as values, but like every function can error. And so like in Rust, like with result, right? You know what I'm saying? Like it has errors as values, but like... you like the language should just tell you what exceptions can happen in something, but you should not have to necessarily, there's like a trade off there.

But anyways, I don't, this isn't, this isn't, I know this is programming. I said, go on top. I know. Well, it's the standard about exception tracking, which by the way, Zig also does that. Zig does a pretty good job with some, is it some level of error tracking? Yeah. Effect mentioned. I'm not sure. Yeah.

I'm not sure if this is kind of the same thing as effect, but I think C++26 is talking about something similar. Yeah. Like they want to have the language also be more expressive of like function contracts, less like function level, more like parameter level. But like you have within the function contract, like, hey, N must be greater than 1024.

Right. You know what I mean? Like being able to guard clauses. Yeah. Like guard clauses, but like built in syntactically to like the pre-processor, not to run time. So it almost like it's like a zig compile time thing. Which I think from a security standpoint is pretty cool. I'm not sure if it has like the same AI effect that you're describing, but same thing, right?

Making the language itself more expressive of like guards, if you will. I would assume every single time you can put in one more like pillar of things you can't screw up, AI just does better. Right? So like if you can say, hey, this thing can't be smaller than this, it's just going to do better. It's just going to go, okay, I can do this.

The corollary for that two prime is anytime you can possibly add a language feature, C++ will do that. And no one will use it by the next generation. Yeah. But agents might use it, Ed. Sure. So that could be different. That could be... Ed, I think you've kind of been sleeping on the fact that C++ is, in fact, the greatest language because agents can write everything.

It has all the things. Yeah, it has everything. Can we get back for a quick second since we're talking about languages and we're already in this rabbit hole for a quick second? Yeah, of course. Is there value... Like this is just a purely hypothetical question here. Let's pretend that agents do the same leap in capability that they did from last year to this year for cybersecurity.

And they do it that same proportionally one more time. What would be the value of something like Rust when you could just be like, agent, make sure there's no bugs. Make sure there's no security vulnerabilities. And it's just like, I am now so good at security vulnerabilities. I will spot every possible problem.

Yeah, I was thinking about this a lot. So earlier this – I mean, fuck, it might have been like earlier this month. I don't know. I'm losing track. But Citrix, right? Citrix is a company that makes a variety of security tools. The oranges. Yes, Citrix. Very close. I love those. I love Citrix acid. That's one of my favorites.

They make oranges and appliance firewalls. That makes sense. And so two bugs came out. We're literally in the authentication for logging in to this firewall. There was a something, something. The N for a mem copy was sourced off the network, right? Bad. Okay. You can't do that. You got to overflow. You got to memory.

Oh, wow. They did a mem copy straight off the network. Yeah, not great. That's actually like a lot, a very common and like embedded devices. It's like, oh, TLV protocol, but the L is user control and then we don't bound it. Or like if we do, we like add a static header size to it and don't check for an int overflow, like crap like that.

Anyway, I know that like embedded stuff is really slow, but it can't be so slow that you can't check like two bytes before. Yeah, it's not a speed thing. I think in the industry, it's like the embedded people get the device to boot. They get the like the the voltage regulator to regulate the voltage.

Right. So like once they're writing the code, it's less about security. It's more about performance. Like think about it. It's an application gateway. So they're. optimizing for all the traffic has to shoot through the thing at light speed and only divert packets that matter, right? That's where their heads are at.

So I don't think it's that they are worried about the speed of a mem copy checking for the N. I think they just don't think about it, right? And again, I'm speaking largely for an entire industry. I'm not trying to do that, but that's the pattern that I've observed. So in my videos, hardware is just also hardware is just harder than software.

So like it's true. Oh, only I just only have mad respect for those people that the fact that they also happen to have to write software. I'm like, I'll forgive you guys that you forgot the mem copy check. I don't know. You guys are building like the fabric of a router. Yeah. In my videos, I talk about like, would Rust have fixed this?

Right. Would Rust have allowed for this to happen? And the answer is, yeah, Rust would have fixed a lot of these issues. You have like obviously like business logic or like command injection bugs that it wouldn't have fixed. But bugs like that, it would have. But the issue now, to your point, is like, bro, if I had an AI agent write that same code, it would not have made that mistake.

Like flat out. It would never, ever have allowed... What's that? Current gen, no chance. It's not writing the wrong piano. No chance. And so, yeah, the question is, like, does Rust even matter from that perspective, right? I think Rust still matters for a use-after-free scenario, right? Because A lot of use after freeze arise where two separate temporal states of a program can exist at the same time.

And the AI agent may not have a large enough context window or knowledge about the code base to even acknowledge that state A and state B can exist at once where a use after free can occur, like a race condition sort of thing. So where it's more of a question of like... Function contracts, that kind of the tedious point, that can still occur.

But I think spatial bugs, like a mem copy, is Rust necessary there? I don't think so. I think the agents just write better code now, and it's not going to be a thing. And again, I'll probably eat these words, but I don't know, 10 years maybe. So we'll see. That being said, the counterpoint there is Sam Altman has outwardly said, I don't know if it's possible to patch all the bugs, despite literally being the lead of an effort called Patch the Planet.

So I don't know. I have no idea. Can we stop there for a quick second on that? Because I'd actually, what in your opinion does that mean? I'm going to give the layman's approach first, which is in my head, that means there are bugs on my CPU inside of its programming that can't be patched. that anybody could just take advantage of if they can construct the correct program and screw me, like to the point where we can't stop certain execution patterns from happening.

And thus we're like DOA on an entire generation of devices. I don't know what he's saying there. I think what he's saying is that the rate of code churn is still greater than the rate of fix. right? Like, there's more code coming out, human-written likely, or even maybe some bad AI-generated code. Definitely some bad AI-generated code.

Yeah, definitely. Just check out my repos, it's bad AI-generated code. I think he's saying that the rate at which we patch is less than the rate at which we code. So, like, just from a pure in-and-out, you know, bucketful of water with a hole in it perspective, we're gonna have more bugs than we have fixes.

That's his perspective. Um... If your question is about, like, are computers... Our computer is designed in a way where security is even possible. There are hardware mitigations that have been out for a long time. Granted, no one uses them. We have control flow integrity, where literally you cannot return to a function that does not have this specific marker on it.

A lot of new code, like Mac OS, is notorious for being the most secure code because their code bases rely on all of the modern mitigations. They have memory tagging, they have control flow guard, they use pointer authentication, all this stuff. We are trending towards hardware. Do you use Work OS to authenticate the pointers?

How does that work? That's a good ad integration there. Whoa, whoa, whoa. What is happening right now? I'm sponsored by Work OS. Yeah, that's really good. Put it in, Ed. Thank you. Wait, actually? We do have some of them. Yes. We've done several of them. That's really good. Thank you. I forgot where I was.

Free for your first million pointers? We're trending towards... Free for your first million. Wait, what the fuck is happening? I have no idea. I'm just making Prime. I don't know. Prime went there, though. I love that. I love that. That's a good one. Thank you. Sorry. It's a good one. Keep on going. Dude, I can tell right now there's going to be a wall of YouTube comments being like, this episode was good except for all the interruptions.

Oh, my God. You know that is going to be the number one. No, it's all good. I wish Deitch didn't talk about his language so much. I get to come out of my TISM rabbit hole, and then I lose track. It's all good. Anyway, yeah, man. So I think computers are becoming physically capable of having security.

It's just we have to use all the features and not write garbage code. I guess that's kind of how it's been forever. But, you know, here we are. That makes sense. Yeah. Yeah. It does feel like there's probably some, there's also just like a subset of, I just didn't like thinking about hospitals and stuff where it's like, they're still running windows XP and you're like.

I don't know what else to say Look at like We have a guest wifi network And you're like this hospital is going to be toast Here in the next few months Look at Minnesota right They got hacked supposedly by Iran And like how did that happen Oh they left like the debug interface For their PLCs on the internet And oh Rockwell Automation Put out a CVE for those same PLCs Like a couple months beforehand They didn't patch And so what was funny is there was an article that said, like, yeah, Anthropic deploying mythos to the state of Minnesota to help secure their power system or water system or whatever.

And it's like, wow, I can't wait for them to spend a million dollars on tokens to be like, take it off the Internet. You know, I think that the positive, the net positive, I think, is that. AI has made cybersecurity kind of a hot topic. I think people are becoming more aware of this, right? You know, it's not going to stop people from leaving shit on the internet that they shouldn't leave on the internet.

Like, look at like the fridge thing too, right? The people had their fridge controllers for like the country of France on the internet. So hackers found a bug in that because I think the new assumption is that Even if there is an authentication interface for a password and username are required, there are likely still bugs in that thing.

So the question is like, not if, but when you get hacked. And as a result, just pull it off the Internet. Put it behind a tailscale network. Put it behind a Palo Alto firewall, VPN concentrator, something that just like at least increases the complexity of attack, I guess. Yeah. I was talking with my local bank about just the security practices they do.

I was just like, well, you know, there's like a lot of dangers. I mean, you have a public website up, you know, and you just got to be careful because, you know, with all these things happening, I think there's like a high chance that we're going to see some news, some piece of news in the next few weeks of a bank having some sort of major security breach.

I assume people deploying these things aren't just going to be like, we're going to hack fridges, you know, like I think they're going to probably go for something a little bit more profitable to figure out. And the, the person I was talking to is like, well, there's no way it's in this separate system.

It's in all this kind of stuff. I'm like, well, just... There's like these ideas of lateral movement and stuff you can't expect. Like just because you say it won't happen doesn't mean it's not going to happen. Just because it's in a separate bucket doesn't mean it's not going to happen. I saw a poll on Twitter and the poll was like, what does air gapped mean to you?

And the two options were physically disconnected from the rest of the world or on a separate VLAN or something. And like 30% of people voted VLAN. And it's like, okay, so we now are acknowledging that a lot of these networks that are truly separate or they're not connected, they're behind a firewall that maybe hasn't seen an update in eight years.

So it's like... Well, I mean, again, October 23rd, OpenAI model escaped in its offline sandbox. Yeah. Escaped where? I think that was a DNS issue. It went and hacked something. Well, it went somewhere and hacked the outside world from its offline sandbox. So it was, in fact, not air gapped in the sense that there was someone's doing it.

it's insane dude yeah it is crazy how much um escape the sandbox has occurred though with these models like they are they are really good at finding vulnerabilities like what is it jfrog artifactory right yeah like it's just like oops okay we're using artifactory so we're gonna have someone on the outside push through a malicious ruby shell ruby rails you know repo that'll pop a shell on there and then we can just get out cool like insane Crazy shit.

I have a question for you, Ed. Yeah. My assumption is, like, these hackers are just running models 24-7 to, like, find things. Do we just, on the defensive side, do they just run models 24-7? They're just, like... Whoever's model's better or something, who has a top-secret model can beat the other model?

Right. How does that work? So I was talking to the chief customer officer at Sentinel-1. Sentinel-1 is an EDR company, endpoint detection and response. And he actually was saying from the telemetry that they're seeing – All that's really increasing is like kind of the speed and scale that actors can operate.

But there isn't like any new technology, right? He called it the cyber physics, which I love. Like the physics of the domain aren't changing, right? You still need an initial access vector. You still need to persist. You still need to laterally move and privilege escalate, right? So what is an offender going to do?

I think an offender is actually going to operate the exact same way. They're still going to look for IOCs. They're still going to look for... malware in their networks via behavioral or holistic analysis or signature based analysis. And they're just going to use AI to kind of automate, automate's the wrong word there, but like, you know, go like run their workflows for them.

Right. But I don't think from a defender standpoint, a ton really has changed. What's his perspective? I've, I'm not a defense guy. I've been a security researcher, so I can't speak too intelligently to it. But I think the media is pushing this whole kind of like we have AI swarms and like this swarm over here is fighting this one over here like a dogfight in AI space.

I don't think that's the case. I think a lot hasn't changed. It's more just the vulnerabilities that we're finding. The number is going up. But my buddy, my podcast partner, Matt, he actually said that he saw a graph where. We're seeing the number of CVEs go hyperbolic, parabolic, whatever, math numbers.

I can find that. Hockey stick. It's going big. Number go up. Straight up. But the number of CVEs that are used to actually compromise, that are used in active attacks, is a flat line. And it's the same CVEs. Nothing is really changing. So it's to your point to you, just more of like less of a vulnerability problem, more of a update your shit and take it offline problem, which is kind of interesting.

Yeah, I do think one of the things that people are just not I was just talking to somebody trying to explain to them, like, why are these like hacks happening like to a non-technical friend? I was just talking to them about it. And I was like, one of the things about it is just like you have to think about as with any other system, you have to think about the economics of it, right?

And before there's first, not a lot of people just like want to get up in the morning and ruin other people's days as the way that they're making a job. So that already makes like, we're very fortunate that on average, that's not what people desire to do. Right? Like, and so that's good. We're like very happy that most people don't want to take hospitals offline.

They think like, I'm willing to hack, but that's scummy. You're like, that's good. That's so that's good. So that like already makes it more expensive. Right. Because you have to find people to like who are actually willing, willing to do that. Right. And then there's like all these other things where, you know, it seems to your point, it does not seem as of today that there's this mysterious group of LLMs buying themselves compute to go hack places to run themselves to perpetuate themselves.

Right. Sure, maybe that is something that could happen in the future, kind of a separate discussion. But right now, it's like, cool. So before, someone could go ahead and do one hack attempt a week, and now they can take the thing that they knew about doing one hack attempt a week, and they can make 100 or 1,000 of them.

Right? And, like, so that just makes it cheaper. It's not necessarily so much that, like... They're like magically coming up with these or anything else. Right. It's more about, I think a lot of it is you just think of nice problem. I don't know what you're doing over there. I don't know what's going on either.

I'm being hacked. I'm being hacked. A lot of it is just right. It's cheaper to do them than it was before. That's really the thing. And so when something gets cheaper, you get more of it just like any other system guys. Like that's the, that is just a fact of life. In fact, Well, kind of, but not exactly.

It's opposite side usually. Yeah, right. Is that one of the paradoxes from the rules episode you guys had or is that a real one? No, that is a Jevons is a good one, but it is not. We're just talking law of supply and demand mostly, right? That when something gets cheaper, it moves on the demand curve and then more people are willing to purchase something at that price, right?

Like foreign actors would like to have every exploit. But they are not willing to spend $100 trillion on having every exploit, right? So but there is a curve where those two things meet, where some amount of exploits and some amount of money, they match. And they say, it does make sense for us to spend $50 million to find this many exploits, right?

And if suddenly exploits get cheaper to get, then you're going to see more of them, right? Yeah, we saw that with the Vercel million-dollar sandbox challenge, right? Like that guy found a KVM escape. That is a bug that like on the black market back in the day was like a $2 million bug. He got 50 grand for it.

And there was a whole argument about how like he kind of got robbed on that. He did it for the love of the game, whatever. He probably could have gotten like closer to $1 to $200K for like a proper bug binding program. So that is way cheaper than it has been for the last five or so years. So to your point, supply is going up.

What do you mean by black market? A bug on the black market? What does that mean? So when you like find vulnerabilities. Trash, first off, Ed, before you go, Trash, don't get any ideas. Yeah. Now Ed, yeah. So where, where can I find this black market? The thing is just asking for a friend, how would I get on this black market?

I also like, no, actually I think Trash is more like, do they also have Pokemon cards there? What other exploits are in this black market? Okay, so when you're doing, call it bug bounty or vulnerability research, you have your traditional bug bounty platforms where the company that has the software is paying you to find bugs.

It's like outsourced security contracting effectively. They pay you for the bug, you report it, NDA, whatever. There are platforms like Xerodium, Crowdfence, Iris C2, there's a few more that I can't name and I can't think of right now, that they are like a gray to black market where they will buy it off of you.

You really don't know who they're selling it to, likely governments around the world, but they'll give you better than the bug bounty will. So depending on your morality, where you lean on the spy or don't spy on journalists thing. They'll give you 10x whatever the bug bounty would give you. So the market, I think, if you go to CrowdFence right now, a Hyper-V Escape, the price I have listed is like $1 to $2 million.

Microsoft will only pay you $100,000 for that. So your call. You know what I mean? That's just the nature of the game. trash is like i'm suddenly having an interest can we uh can we put those links in the description sorry by the way josh side note take out my hiccups i can't help it right now hell yeah keep them in zoom in on prime every time he does a hiccup please thank you But yeah, man, so that's kind of the nature of the game.

Yeah, so to your point, Teej, like supply and demand. The supply is going up, demand is maintaining, so the cost is going down. Right. Yeah, and also it just is literally cheaper to do because one person who's good at hacking can now do more things, right? Even if it's, like, I still, like, it seems it's possible that some people are still finding, I think, bugs just by literally being, like...

Fable, look at this repo. And giving no advice and other stuff like that. Find me a bug, make no mistakes. Find me a bug, make no mistakes. And it probably does sometimes, just literally through models getting smarter and random chance that it happens to look up a thing that does... that time find a new one.

But like a lot of these I'm imagining are like not just people spinning with no direction or it's like enabling people who know what they're doing to check a thousand places where they used to only be able to check one. Right. Yeah. Like, so my big example for this is like we have had security tooling for a long time.

Like when I was at Microsoft, I wrote fuzzers for Microsoft. What is a fuzzer? A fuzzer is a tool that is optimized for code coverage. It traverses the call graph of a program, tries to enumerate every block in that code graph, and then it mutates all the edges, right? Tries to find weird edge cases.

Cool. What that is not good at doing is finding a use after free. Because like I said before, those are spatial bugs. A temporal bug is one where like two separate states have to exist that kind of collide with each other and enable a weird temporal issue. What's what AI is really good at is like if you can describe the pattern of the code base that could enable a use after free to the LLM, it's very good at automating pattern finding.

And so PHP to shell was actually found by a researcher at Caliph using an LLM where he described the pattern of a use after free. found it in a newer version of PHP, and then kind of mutated. He changed his skill file a little bit and said, okay, look for this less specific pattern across the PHP code base, and in a couple of days found the same bug that's been in there since 2004.

Yeah, so researchers that are already good and know the proper way to describe the bug and the code smell of the code base are just being completely enabled to do 100 times better work. Both for good and for evil. Yes. Yes. Yeah, for sure. So one of the common questions I see is, is this the part where things get really bad?

And then over the next maybe six months, things actually get less bad. And then we kind of actually return, we go to a place where everything's fantastic. And now a lot of stuff is much more secure. Or are we heading into a direction where we just get more and more vulnerabilities? That's a good question.

I don't... I don't know how I feel about this because part of me. No, I need a hot answer now. I need to type. Go, go, go. I need an exact time frame. How many months until things are saved? Three years. No. I'm looking for a 250K view clip at least on X.com. At minimum. Otherwise, not uploading. Okay.

Listen, dude, like, again, this is kind of a hot take, but the LMs are better than most programmers. Sorry. Like, just in terms of their ability to write code that doesn't have spatial bugs and the speed at which they produce that code. So if you look at it that in a vacuum, code gets more secure, in my opinion.

Yeah. But at the same time, we now are in this weird state where anyone can make anything and at any scale, by the way. So you have this weird code technical security debt issue where, like, yeah, the code is more secure, but you have a million lines of it that you don't know what half of it does and you have to maintain it.

So, like, what does that mean for your project in five years? So I think in the short term, we are going to see a lot more bugs. I think we're going to see people getting really good at finding bugs really fast. And hackers are going to hack them a lot quicker. I think in five years, ideally, we're really going to shake all the bugs out of Linux, for example.

Like, we're finding LPE after LPE after LPE. I don't think that lasts. LPE stands for? Sorry, local privilege escalation, right? Look at, like, dirty frag, fragnesia, all of these new kind of, like, page cache corruption bugs. I don't think that lasts forever. I think there are probably plenty more to be found, but I don't think that there are, I mean, there are not infinite, right?

There are only a finite number of bugs. We will shake them all out, ideally. I think that means in the future, code is, I mean, not that code is solved, right? Like I think there will always be bugs. I just think right now to Teej's point, bugs are getting cheap because there are so many. I think bugs go back to being expensive again in, I don't want to give a number, but maybe five to 10 years where we're like, you know, The harnesses for finding bugs are very good.

The cogeneration bugs are very good. And the remaining bugs are just like really weird, obscure edge cases where like subsystem A and subsystem B like got confused on their contracts and like disagreed about something. And that causes a bug sort of thing. But I could be wrong. I could be wrong. This could just be it forever.

We have infinite bugs until we all die. I have no idea. We're talking more specifically for, though, probably the kinds of things like, you know, actual like memory corruption. Like not as because this is one of the things that people are confused about when we talk about this, because it's like there's a whole separate world where it's like, well, you published an app and you didn't put admin auth on anything.

And you're like, right. And like those, that will be interesting to see too. Cause to your point, I think, um, and, and this is like, I don't really know exactly how this is going to play out. And I'm interested to see as well is I think a lot of people who have actually spent a lot of time programming, they know the right things to ask for, to make sure that the LLM like, doesn't do something stupid for them for something like, um, I can imagine this is an easy scenario for me to imagine.

I'm new to programming. I want to spin up a website. And when I'm testing it locally, I say I want it to be able to skip username and password so that my, my agents can test it for me easily. Right. That totally makes sense. That's exactly like if I had like a, just a random lay person friend say they wanted to make an app and they're like, well, every time the agent tries to do something, I have to go and type in the password or I have to like paste it in somewhere.

I was like, that's stupid. I hate this. TJ told me the computer can do everything. So I just tell it, don't require that. Okay. Well now a human told the LM and pushed it in a direction that's inherently insecure. Right. Whereas like I would say, knowing program, be like, cool, let's make a way to mint local tokens that we can do and like have an actual auth system that makes sense.

And then I'll have like a bearer token that's, that I put into a local end file that only works for my dev environment and doesn't work for production. And like, but they're not going to come up with that by themselves. Sometimes the LLM might shoot. Like, I think LMs are choosing more often better options, but like people actively choose bad options, like intentionally, you know, they just don't realize that they're bad.

So there there's this, like, I don't know what's going to happen with like, Oh, random CFO starts automating his workflow. Right. For his company. And then now suddenly the bank account is hooked up to a random endpoint that's also connected to the internet. Yeah. Right. Right. I'm with you. I'm fairly myopic on the...

like memory corruption vulnerability side. But yeah, to your point, I think those kinds of bugs were like, person didn't know what they were doing for sure. You are always going to have some form of like business logic bug where like, it's truly just like a logic issue where they didn't consider like state A and state B colliding or like subsystem A or B colliding.

And then third, yeah, like social engineering and supply chain, like call it like the layer eight vulnerabilities. Those are never going away. Right. Like NPM is making it so that like post install scripts don't run by default. And like we're doing better about like app granular or API key granularity and stuff like that.

But someone's going to call grandma and say they're Ben Affleck. And like there's really nothing you can do about that. You know what I mean? Like that's just going to go on. He knows about AI, so I would trust him. Yeah, right. Exactly. The problem is that if I find a USB on the ground, I do just want to put it in a computer.

He's plugging that bad boy in. He's plugging that bad boy in. You know it's true. It's hard not to. It's not as bad as it used to be, bro. You'd probably be fine, honestly. You heard it here first. Everybody, mail crime. That'd be a fun video, actually. Anyway, it doesn't matter. Yeah, man. The amount of targeting it would take to get a USB stick given to Prime where they know his setup and then exploit while it plugs in is like, that'd be impressive.

I would watch that. I'd go drive by and just throw a bunch of them on his driveway. Yeah. I'd be like, dude, there's so many USB sticks. One of them has to have a secret in it. Let's go. There's a Bitcoin in one of those, for sure. For sure. One of the Bitcoin locked in, for sure. How I think about bugs is actually from an Edgar Diestra quote, which is that debugging is the process of removing software bugs, then programming is the process of adding in bugs.

Yeah. So while we'll be programming, we may very well always be with bugs. Yeah. But again, like, is the... is the model going to write insecure code in five years, in a year? You know what I mean? Like, look at how far they progressed. Because, like, what happened, right? Like, Mythos was good at writing exploits and finding bugs because it got very good at coding, right?

That was the whole side effect was like, oh, I know what bug patterns look like, so I don't make them, comma, therefore, I can find them. And if they keep tuning models to be good at that kind of thing, does it ever write insecure code? I don't know. And this may sound like very copium or hopium or whatever.

It's just like the CVE number is going kind of a number go up. I would imagine code security will go that way eventually. I just don't know when that starts. If it hasn't already, I don't know. If, like, our hypothesis is that it's getting cheaper to find these, right? I mean, which I think is, like, seems pretty widely agreed upon, even by people who are not AI maxis, right?

They're saying, like, hey, they're finding. So, okay, well, that means also if it's easier for, if it's cheaper to find them, that means me as the producer of my own software, I'm also able to find them for cheaper. Right. That demand curve is still different. Like I may not be able to afford to pay a million dollars to find a bug on my software.

Right. But I might be willing to pay like five dollars. Yeah. Right. You know, like I might be I might be willing to pay that for like the LM or for some like research consultancy. That's really good at finding these to come in and audit before we do a release or, you know, like there's there's interesting things that I think we're just like people are very stuck.

Yeah. in like thinking that it's going to be exactly the same, like five years ago as it will be in five years. And like, I think like probably there's going to be different styles of like people finding these bugs where you can do it and you don't have to be like a, you know, hoping that black hat hackers don't find it first and sell it on a market instead of your bug bounty thing.

It's just like, cool. We have like, as part of our release process, either like automated things or people who are really good at this, who can now, Like, Ed, if you were working at a random SaaS company, you would be able to check a lot more stuff than you would have been able to check before. Correct.

I mean, Casey, sorry. So, yeah, on a game engine, you'd be able to find a lot more bugs than you would have been able to before. So I think there is something, too, there that, like... When we think about it in just like regular econ terms, like a company is willing to buy more of it. It's willing to buy more bug protection when it's cheaper per bug.

Yeah, 100%. You know, yeah, scans are getting cheaper. And then, too, it's important to understand, like, culturally, this is also, like, the hot topic. Like, it's becoming more of a... Embarrassment is the wrong word, but like you're kind of expected now to have a security program and you're kind of expected to have like an AI enabled security program and doing that is cheaper than ever.

So I think people are just more opting into that, which is good. Like it's just something that's been I've been thinking about, but this is it's like security adjacent, but I don't maybe you'll have some opinions on it, which was like. Do you, is there any inkling that like, as we either find more like, uh, like crazy bugs in software, or we just like solve a bunch of other comp sci questions.

Like one of the questions says like, will cryptography be dead? Will it be, will we not be able to like solve, will we not be able to have this lopsided thing where it's very easy to check, but very hard to solve certain questions and like, what? That's just been something I've been thinking about recently.

I don't know if you have any thoughts on it. It's also okay if you're like, I have no thoughts about that. Yeah, I have been thinking about this too. Yeah, because if people don't know, right, like public key cryptography is a system where Teej and I can exchange keys. And the backing of that is a hard math problem.

like yeah factoring integer factorization yeah like factoring a large number is hard or like knowing the points on this elliptic curve i don't know elliptic curve math but like generally something something curves hard um but if ai is able to just like one shot the hard problem and like find a way to reduce the entropy of rsa by 10 or like 10 bits or whatever yeah i don't know where that goes i don't really think i know enough about like the math fundamentals like like do math hard problems actually exist or are we just too dumb for them you know what i'm saying like right yeah that's that's part of what i'm kind of wondering or like the literal hypothesis of the exponential time hypothesis which is that it's so hard to be able to do the three sat thing that it can never be better than um exponential and that's why open ai's big kind of publish of hey we actually found sub exponential was so important because if they prove that they can actually

And if it can keep on going down and we can eventually get into polynomial, all bets are off. Right. Yeah. Right. So what does that mean long term? Potentially the public key goes away. I mean, there's like the whole world of quantum that I'm not going to pretend to understand. Like maybe there's quantum crypto.

I thought that was elliptical curve was supposed to be like, oh, it's quantum secure because it's like. Oh, yeah, because like it relies on past state and there's like all these kind of rules you have to have for quantum to make it. So, yeah, because the main vulnerability to public key was that RSA could be way, way more easily factored via quantum using like Shores algorithm or some shit.

I'm not going to pretend to do the math. But yeah, I think you're right. I think elliptic curve, Diffie-Hellman, is not broken by quantum. But yeah, that's assuming that AI doesn't find a vulnerability in the underlying math of elliptic curve. So, yeah, man, I don't know where it goes. It is an interesting question because it does kind of put into question the underlying architecture of the Internet full-time.

That's how everything runs. I was just interested if that was kind of like a thing going on getting discussed in the security community or like you guys don't have time to talk about anything besides the. No, I saw I saw a respected cryptographer on Twitter. I think it was like Matthew Green. He was like, yo, public.

He might be cooked. And it was a link to one of the open AI or entropic findings on like the polynomial thing you're talking about. Yeah, which would be crazy. Like people don't really understand. I think that like cool, like banking is online. Banking is done. You know, it's like, like we don't, you know, like encryption is like encryption as we currently do.

It is done at that point. Like. Yes, because if that was broken, it would also break like AES, if I'm not mistaken, SHA-256HMAC. There's all these, and also Bitcoin, which is awesome. But it'd like break a bunch of stuff where you actually couldn't. Well, all the symmetric encryption we do between like on a channel on the internet, that key exchange is done over Diffie-Hellman.

So it's like if you break the Diffie-Hellman by breaking whatever math is underlying the exchange you're doing, like that's now you have the symmetric key. Now you can encrypt all the traffic. By the way, for those that are wondering, some mathematician posted that he thought P equals NP was like a 0.1% chance to be true.

And then after yesterday's paper, he said it's a 2% chance to be true. And so like that, I mean, that's a terrifying amount because we, even though like P equals NP actually happening, it would make like curing cancer and all these other things really, really easy. It would also make the internet impossible.

So there's, there's a lot, it would really change our, our world quite a bit. Yeah. Yeah. People might actually find out what air gapping is. That would be great. Actually strange times, man. Trash, you got any? I feel like, Trash, I don't know if it's just because you've had this mask on. Maybe you've been talking and we couldn't hear you.

But, like, you got any other questions or thoughts? I have a fan under this mask because I'm so hot. A fan? Nice, dude. I got a neck fan because it is so hot in here. And I'm just dying. But the thought I've been thinking about this entire time since this has been going on was, like, What happened to all these people that claim they're like security experts and then all their shit just blew up?

Are they just like Homer Simpson meme just like into the bushes? And then like, that's that. Like, has everyone just been LARPing? Like, what's going on here? What is there a specific instance? Like what security experts got blown up? No, no, I'm just saying, like, in general. So it's like, I'm just, like, some expert that got hired at some company to write some super secure code.

And then AI come and just blows it all up. And I'm like, oh. Yeah, I don't think it's a LARP thing, dude. I think it's just so easy to make a mistake that really blows up everything you... Like, it would just take... a attacker to have the right exploit and know my phone number to like get on my phone and nothing to do with me that's just like the the physics of the domain are like stacked against me right i think that applies to any any any security consultant that's been caught with their pants down right it's like shit just happens and it's why are their pants down when we're talking about security Well, you know, some people are really into security.

It's actually a problem. So, yeah. I would say with that, too, I mean, at least this is the discussion that I thought, like, one of the discussions I find really interesting about it, too, is just... When you're on defense, you have to get everything right. But when you're on offense, you only have to find one thing that's wrong, right?

Which is sort of like the, I think this is sort of the contrast that a lot of people are discussing in the security world of like, cool, maybe we get much, much better at writing, you know, exploit-free code. But, like, if that number is 0.00001% or something, and also we're generating 10 trillion times as much code.

Right. What is the... Does it end up... It might be that we're, like, way... You can think of it as, like, way more secure in the sense of, like, I used to have... one out of 10 lines of code was bad. And now I have one out of a billion. Okay. But if we get way better at finding the code or we get right billions of times for code or something, then we can end up like less secure.

So that's one of the things that at least from just the outside, looking in, watching that discussion has been interesting to consider there. I don't know if you have any thoughts about that, Ed. Yeah, we talk about this a lot. People tend to say that cyber is asymmetric in that attackers have to be right one time.

Defenders have to be right every time. So my buddy Matt, same thing. He actually hates this discussion. He is more of a defender guy. He's been on the blue team his entire life. And he says that... In this world where everyone has AI powered everything, so do the defenders, right? Which makes sense. Like, I agree with that.

Um, and so I actually talked to the lead of Sentinel One's, uh, threat intelligence labs about this last week at their conference. And he said, like, he's actually seen a bump where people that are doing defending him included can now write like, um. insane heuristic and behavioral analysis tools on the fly with like by vibe coding that they never could have in the past right that like you know the attackers have their their exploits but now defenders also have these really cool tools that they can just spin up basically for free um so i still i think i agree that it's an asymmetric um like in conflict if you will but defenders do have an advantage there too which is kind of cool to see So home turf is really starting to pay off now.

Yeah. I would assume that's the case because, again, it goes back to the same kind of rust question I have, which is once they get good enough, there has to come a point where it just becomes increasingly hard to do stuff and where defenders actually start meeting where offenders are. I agree. I agree.

Yeah. I think it trends towards things being more secure. It's just going to be a very painful end number of years to get there, unfortunately. But we'll see where it goes. I got one more question unless someone else has one too. But just like thinking about some of the like language discussion from earlier and stuff like that.

I think one of the things I've been thinking about too is like, let's say you're going to be running a lot of like LLM review stuff automatically and all these other things. Like, is there a discussion in the security world right now about figuring out how to either like through languages or through tests or through specs or like something provide like more concise?

I literally mean like more concise, like token efficient ways of describing things about your system and like invariants and stuff like that. Because literally like... the token efficiency of something will greatly affect the bottom line of how much your ai spend goes by the end of the year right like if you have to fill up your whole code base into into context every time like let's say we get a model this year that has 10 times the context that doesn't mean you want to fill it up like that's the you know it's like the token efficiency of things like actually matters is that like a discussion at all going on in the security world So this is what people haven't thrown the word harness around a lot.

Oh, yeah. In the security world, this is what people mean when they say security harness, right? And I don't have any like concrete examples of this like that I can talk about. There's one I worked on back at like Microsoft or whatever, and the others are obviously like, you know, Cloudflare proprietary and like Amazon, whatever.

But the conversation is about like exactly that. How do you make it A, the most effective and B, the cheapest? A lot of the sauce is like not really even using AI to do the bug finding. It's having the AI write tools that are non-inference based. Like, hey, AI, go write me a harness that will be a fuzzer for this code path.

And then at that point, you're not on inference time. You're on compute time. Stuff like that. And just really tuning in. that process to reduce the overall AI load, right? Because again, if you're looking for like, use after free type bugs that are not good at being fuzz, like sure use AI inference to find those patterns and stuff, but it's still on you as the harness author to like, don't give it the whole code base, like only give it, you know, source code A and source code B.

And then even then maybe you cut it down to like those specific pattern that you're looking for and reduce overall loads. Yeah, that is a conversation for sure. I don't think that any one person, any one company has the best tool. I think it's very tailored to like the kind of bugs and the kind of code base you're working on.

But that definitely is a conversation. Less popular or less open, I guess, than you would think that it should be. But yeah, people are talking about it for sure. Yeah. One last one, last one. So I know people that are probably listening to this are going to think, like, is my code insecure? And I'm not an expert on this, but I know there's security-specific models.

How do you go about getting access to one of those? Do you have to be working for an actual security firm? Or how does that work? So I don't know the full process. I know for Anthropic, you can go try to get accepted into their cyber verification program. I think before Mythos was in it, I was in it, and now I'm not sure if I qualify, but whatever.

It's just an application. You have to show your passport and show you're a real person, work for a real company. Not that complicated. But there are other ones like Daybreak Astra and Daybreak Blue or whatever that I really don't know. It is, unfortunately, and I could be wrong here. Sorry if I am. Kind of a good boys club where you got to kind of know a guy to know a guy.

Look at DAX, right? OpenCode. You have to get hacked by OpenAI typically to get in. Like literally, Dax had to get hacked before he was allowed to touch a security model. Like that is not the right answer, in my opinion. So if you are concerned today, I would argue use, I think Mythos 5.1 is public now.

And then GLM 5.3 are also good if you get restrictions on that. They're pretty good at finding bugs. They're also a variety of... AI SaaS tools that are out there, static analysis security tooling, many of which sponsor my channel, but I'm not going to do that plug here. But they do a good job. Like they found bugs in the code that I use for my production apps and for my personal stuff.

So they're very good at taking traditional SaaS knowledge, like static analysis, security or composition analysis, and then wiring AI into that process. So all of those tools exist. But are you going to get data big blue? Probably not. Probably not. Any thoughts on people running local or open models or like even renting or like are there providers out there doing like, hey, you can use this model for whatever the heck you want?

Like rent, you know, is that like a thing yet or still not really a thing? Yeah, I mean, so I personally would like to get to a point where I buy a rig and just do, like, agentic loops forever. But I don't want to spend the $20,000 it would take to do that. So right now, I run GLM-5.3 on OLAMA. GLM-5.3 is obviously a Chinese model.

So if you have, like, concerns about sensitivity with national security stuff, whatever, like, maybe you don't use it on ZAI's hosting. But, like, OLAMA hosts it on U.S. and U.K. infrastructure. So, like, if you're concerned about that, that is a problem. Are you going to be quantized or no? What's that?

Is it quantized or no? I think it's full. You can run... They also have Flash on the cloud, but the one that I run is GLM53. I'm pretty sure it's full, non-quantized. Same thing with DeepSeek. I think Alama also runs DeepSeek. I'm all for the local thing. I think, honestly, people should try to figure out how to run models locally because I'm very concerned that, like...

Sam Altman and Dario and friends are going to like get open weight models banned and like make them the national like intelligence provider. And I don't like that. I want to have my, my computer be at my house and me be in charge of my compute. But obviously not cheap, especially now we're like a 50 90 is like a fucking, what, like $7,000 graphics card.

That's not affordable to most people. Yeah. So something that's pretty interesting, which is I was talking with Dax after the open code thing. And I, I asked him like, how much is he actually spending on security now? And he's, and he gave me a number that equated to approximately a million dollars a year now in just inference of just like security inference, just grinding away and preventing himself.

Cause obviously if you're storing keys, if you're storing things that people want to steal, you're going to get attacked way, way more compared comparatively to other, you know, other stuff. But I did want to, um, I did want to kind of shut down this old podcast, call this an episode, but I really wanted you to have the last word.

Sorry, I keep, I still have hiccups. I still wanted you to have the last word. And so I figured the best way to do that is to give your prediction on when the next 10.0 CVE for PHP will drop. For PHP? For PHP specifically. Ooh, ooh, ooh. Great question. ETA. I'm going to say two weeks, man. Okay. Two weeks.

All right. Because here's the thing. I talked about this a lot in my video last week. The entire world, the entire world pointed an AI model, very competent ones, GLM-53, DeepSafe-4-1, probably some people with Daybreak access, right? Everyone pointed their models at PHP to try to find the bug before the patch came out.

Which means I'm quite confident that somebody else found a different bug in PHP during that time. And don't forget, the bugs that Kalief found were found in August, right? And this was them working through disclosure, working through patches. So I would not be surprised if in that process people found other bugs.

So, I mean, if the answer, if the question is like, if someone has found a bug, it could be right now. But I think like within a couple of weeks, maybe two weeks, we'll hear about another 10.0, 9.2 for sure. All right. All right. Thanks, Casey. Thanks for having me. Appreciate it, guys. Legendary security researcher, Casey Mortari.

Author of Vulcan. Let's see. By the way, Ed, where can people find you? I'm on the tubes at low level TV. I'm on Twitter at, at low level tweets. I'm also on Instagram at low level TV. And I, I stream sometimes, but not really lately. I might be streaming on YouTube live more, but anyway, that's, that's me.

Yeah. Awesome. Cool. Thank you very much. I hope that you enjoyed this standup. Please next time come with your blockers. Okay. I have not been seeing a lot of blockers, but I'm not seeing a lot of progress. I'm not seeing a lot of code. So come on. We need to see what's going on around here. Sorry. I spent all week planning on my costume.

So that's it. That's my job. It's a sort of Michael Scott situation. All right. Also, I love that Trash wore pure silicon on his face and is shocked that he's sweating. Like, I just love this idea that he's just completely blown away. The fan? The fan is epic. The fan is saving me. Why am I so hot? Crazy.

All right. I got you. I blew up the day. Vibe cold and errors on my screen. Terminal coffee and handbuff.

Security Is Getting SCARY

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin