# The Mythos Problem: When AI Finds Every Vulnerability

URL: https://securitytable.ai/the-mythos-problem-when-ai-finds-every-vulnerability/
Published: 2026-04-15
Duration: 2838 seconds
Season: 4
Episode: 8
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

**In this episode, we break down the “AI Vulnerability Storm” and what happens when AI can find—and exploit—vulnerabilities faster than humans can fix them.**

**We explore how compressed OODA loops are shifting the balance toward attackers, why traditional scoring like CVSS may start to break down, and whether “just patch faster” is even realistic anymore. The team also questions the push toward AI agents everywhere—and whether fighting AI with more AI actually solves the problem.**

**At the end of the day, it comes back to fundamentals: reduce your attack surface, simplify your systems, and focus on what actually matters.**

**  
🚀 Join the Conversation  
Is this a real shift in security—or just faster chaos?**

  

Follow AI Security Table:

➜ Home: [https://securitytable.ai/](https://securitytable.ai/)

➜ X: [https://x.com/AISecurityTable](https://x.com/AISecurityTable)

➜ LinkedIn: [https://www.linkedin.com/company/ai-security-table/](https://www.linkedin.com/company/ai-security-table/)

➜ YouTube: [https://www.youtube.com/@AISecurityTable](https://www.youtube.com/@AISecurityTable)

## Transcript

**Izar Tarandach (0:10)**: Just had to look at the microphone.

**Matt Coles (0:12)**: Just look at it and go, ah!

**Chris Romeo (0:15)**: Hey folks, welcome to the Security Table where Isar is looking at his microphone, which I don't think makes it work any better, but it's a place to start. It's a new day. A new day.

**Matt Coles (0:27)**: My, my girlfriend is here. I have, I should be looking here. That's called a, that's called a monocle.

**Chris Romeo (0:34)**: And you could, you could totally rock the monocle if you went for it.

**Matt Coles (0:38)**: I can't take my glasses off cuz otherwise I poke myself in the eye with my monocle.

**Chris Romeo (0:42)**: Well, we, we are a safety-first organization here, so please do not take out an eye with a monocle. All right. Well, the topic of the day, the day is a paper that's just been released that is encapsulating a pretty big issue that's going to impact cybersecurity for pretty much ever from this point forward. And the name of the paper is The AI Vulnerability Storm: Building a Mythos-Ready Security Program. And so between Mythos and Glass— Project Glasswing— Mythos being this new Anthropic model, which is good at finding vulnerabilities, it's a little bit too good at finding vulnerabilities in systems and creating exploits. And then Project Glasswing is the collaborative collective of various organizations coming together to share information under this heading of AI-generated exploits. So that's kind of the setup, I guess, of what I understand about Mythos and Glasswing. What's, uh, what, what is this paper trying to do? Like, what's their goal? Anybody able to figure that out?

**Izar Tarandach (2:06)**: I think that the first big goal here is that they are setting up the next war between G for GIF. Is it mitos, mitos? What, what, what, how did, what?

**Matt Coles (2:19)**: It's mythical.

**Izar Tarandach (2:20)**: It's mythical.

**Matt Coles (2:21)**: Remember the mythical man month?

**Izar Tarandach (2:24)**: Yep.

**Matt Coles (2:24)**: Now this is the mythical LLM day.

**Izar Tarandach (2:27)**: So is it mitos, mitos?

**Chris Romeo (2:33)**: Somebody answer that.

**Matt Coles (2:33)**: Are you Greek or are you?

**Izar Tarandach (2:36)**: Oh, that's a very good point.

**Matt Coles (2:41)**: Uh, you could ask, uh, you could ask, um, Gemini or something to, uh, to pronounce it for you.

**Izar Tarandach (2:46)**: It's just going to laugh at my accent.

**Matt Coles (2:49)**: I know it.

**Chris Romeo (2:49)**: I know it.

**Izar Tarandach (2:50)**: That, that, that they're getting too smart, but actually the, the, the mythos here. Okay. I'm going to go with mythos. So the mythos here, just to keep on the Greek, uh, theme, is setting up the next, uh, Sisyphus effort. If, uh, people are not familiar with my accent or with Sisyphus, it's that guy that is condemned to roll the big rock up the mountain just to see it fall on the other side, and then he has to start again and again and again.

**Matt Coles (3:20)**: Is it better to be Sisyphus or Daedalus?

**Chris Romeo (3:24)**: Oh, wow.

**Izar Tarandach (3:26)**: Or Prometheus, the guy with the eagle pecking on the liver. Yeah, that's—

**Matt Coles (3:31)**: Lots of doom and gloom in the Greek mythology.

**Izar Tarandach (3:32)**: It's like, man, the Greek had a lot of time in their hands. Whoa.

**Matt Coles (3:38)**: All they did was sit around and like philosophize, right?

**Izar Tarandach (3:41)**: Right.

**Matt Coles (3:41)**: Philosophize. I can't say that.

**Chris Romeo (3:43)**: Wow.

**Matt Coles (3:44)**: I'm sure they had linguists too, but.

**Chris Romeo (3:46)**: You got it right the first time.

**Matt Coles (3:48)**: I did?

**Izar Tarandach (3:48)**: Philosophize? Yeah. Yeah.

**Chris Romeo (3:50)**: Philosophize on something.

**Matt Coles (3:51)**: Wow.

**Izar Tarandach (3:52)**: Actually, I could sign up for that job.

**Matt Coles (3:55)**: You'd have to wear a toga.

**Izar Tarandach (3:57)**: Just sit down. Oh, I can rock a toga. But just sit down and philosophize. Sounds good. Okay, let's philosophize a bit here. So, uh, full disclaimer, is our critique. Yeah. I, I, I, I have read up just to the, uh, executive summary and that they were very nice to, to give us a, a point by point, blow by blow.

**Matt Coles (4:23)**: Playing. Well, we should say this is a, this is a paper in draft form.

**Izar Tarandach (4:26)**: Oh yeah.

**Matt Coles (4:27)**: Provided under common something, uh, Non-commercial, common, CC, whatever.

**Izar Tarandach (4:33)**: Uh, oh, come on.

**Matt Coles (4:35)**: By CISA, or by, sorry, by, by, not by CISA, by the former director of CISA.

**Chris Romeo (4:41)**: Cloud Security Alliance. Cloud Security Alliance.

**Izar Tarandach (4:43)**: CSA, CSA.

**Chris Romeo (4:44)**: Collective plus genaisecondurityproject.osp.

**Izar Tarandach (4:47)**: Yeah, I always got this confusion between CSA and CISA. And what a community it is! Because we got names like, uh, Jan Esterly, our beloved Bruce Schneier, Phil Venables. Who else I recognize here?

**Matt Coles (5:04)**: Katie Maceris.

**Izar Tarandach (5:07)**: Katie.

**Matt Coles (5:07)**: Katie. Thank you.

**Izar Tarandach (5:11)**: Who else? Rob Joyce, of course. And Gadi Evron, who's been doing a huge job with AI lately, AI security lately. And Yeah, I'm sorry, but I'm bad at names.

**Matt Coles (5:26)**: Rob Van Den Veer is on here.

**Izar Tarandach (5:28)**: Oh yeah, yeah, yeah, yeah, definitely. But, um, and the whole, the whole team over there in SANS, right?

**Matt Coles (5:34)**: Ed Skoudis and others.

**Izar Tarandach (5:37)**: Yeah, but just, I was telling you, you guys before, just like starting to scroll through the pages, I get to the authors and contributing authors one, and I'm like, whoa, okay, I can't say a single word against this thing because Nope. Like this, this huge brain trust right here on page 2 probably didn't leave anything to be, uh, to be, uh, questioned. So let's close shop and go home. Yeah.

**Chris Romeo (6:07)**: There's always a time to question. Uh, yeah. I mean, I kind of read it, I, I guess I interpret it a little bit differently in that this is likely the existential problem of the intersection of cybersecurity and AI for the next 5 years. So it is getting a lot of attention. It's getting a, a lot of people are thinking about it because we've reached the point where AI can break stuff at a 100x time faster speed than humans can, maybe even more than 100x.

**Izar Tarandach (6:44)**: Mm-hmm. So I'm going to agree with that, but with a slight aside. In the past few years, which as we know in AI technology terms are decades, we have seen many times something hyped up, something declared. And then when we got to look under the hood, it was all of a sudden, ah, it was in these conditions, but not those conditions. And the wind was going in the right direction. The moon was in the right place. Apart from a very select number of people, nobody has actually seen these results and measured and tried and reproduced results and, and things like that.

**Matt Coles (7:28)**: Well, and, and just, and just to highlight it on page 10, it does actually talk about the Mythos experience. Internal testing at Anthropic showed Mythos generated 181 working exploits on against Firefox. Whereas Claude Opus 4.6 only had 2 under the same, same conditions.

**Izar Tarandach (7:52)**: So, what was the opening? What was the opening of that sentence?

**Matt Coles (7:56)**: 100, uh, Mythos generated 181 working exploits.

**Izar Tarandach (8:00)**: No, no, no, the opening, the opening before that.

**Matt Coles (8:02)**: Internal testing at Anthropic.

**Izar Tarandach (8:03)**: Uh-huh, by whom? At Anthropic. Who produces both models?

**Matt Coles (8:08)**: Anthropic.

**Izar Tarandach (8:09)**: Who are we paying when we use both models?

**Matt Coles (8:13)**: What are you getting at here, sir?

**Izar Tarandach (8:15)**: I'm going to need something a bit more independent before I run away with this.

**Matt Coles (8:19)**: Sure. I'll keep looking, but so far, 181 working exploits against a web browser. So you're talking about JavaScript engine and an HTML processor.

**Izar Tarandach (8:36)**: Oh yeah, sure.

**Chris Romeo (8:37)**: And if you go to the timeline, Isar, I mean, there is more data to back up, uh, page 12, uh, um, February 5th, 2026, Anthropic using Claude Opus 4.6 reported 500+ high severity vulnerabilities in open source software.

**Izar Tarandach (8:54)**: Look, I'm not saying that this is bullshit. I'm just saying that there have been instances in the past.

**Chris Romeo (8:59)**: Sure.

**Izar Tarandach (9:00)**: Things got propped up to a certain level only then to fall down from a cliff when people look—

**Chris Romeo (9:06)**: Yeah, but I mean, I think we can all, we can all admit the fact that even if this isn't it, it is coming. It is coming though. It's like, it's, it's, even if this isn't all that they're saying it is, the technology is moving in this direction and within 1 to 2 years, this will be a reality. Where this model will no longer be in preview only? People, company, you know, nation states, other organizations are going to have access to this, and this is going to change the game on the— it's going to drive software security to be resolved, or people are going to go out of business?

**Izar Tarandach (9:44)**: As it is well known by now, I am not a fighter pilot, but I play one on the internet. So The thing is, for years and years and years and forever, people have been looking at that thing called the OODA loop, the Observe, Orient, Decide, Act loop in cybersecurity. And the way that it connects to my joystick is that it came from like from the military, from, from fighter pilots in saying that in a, in a, in a dogfight, The, the pilot with the, the smallest OODA loop, the, the pilot that's capable to observe, orient, decide, and act in the shortest amount of time, is the one that wins the fight because they can react and predict and act much, much, uh, faster than, than their opponents. And we have been seeing the same argument applied to cybersecurity for many, many, many years. And to me, what, what this whole thing now with AI means is that the, the OODA loops are getting, sorry, extremely compressed. Where things get different, and, and I think that this paper brings that up, uh, very well, is the first three phases, the observe, orient, uh, decide, those three, they, they they scale together, both the attacker and the defender. Using AI, they can make those things go faster. Where things get bogged for me is the act. Because for the attacker, acting is releasing the attack, pointing it, firing it, and out it goes. For the defender, the act is patch those things before the attack comes in.

**Matt Coles (11:34)**: And again— But you're assuming they patch, you're assuming the patch exists. Yes, that you're assuming the patch works. And you're assuming that there isn't a bypass.

**Izar Tarandach (11:43)**: Exactly. Thank you. Which are all amplifiers of that timescale for the defender, right? So I was very surprised when I saw in the page.

**Matt Coles (11:55)**: Oh, I know, by the way, you hope the patch doesn't break something else. Who? Because you have that too.

**Izar Tarandach (12:00)**: And you hope that the IT people who are not the security people who actually say, yeah, okay, we're going to put patch in place. Don't have their own processes that take woodles of time on top of that.

**Matt Coles (12:11)**: Woodles of time for the OODA?

**Izar Tarandach (12:12)**: Woodles. Woodles for the OODA.

**Chris Romeo (12:13)**: A lot of OODA this morning.

**Matt Coles (12:16)**: A lot of OODA.

**Izar Tarandach (12:18)**: Yeah. Turn it into fun. Okay. So the, the, the thing is, uh, so the paper says that what to do now to deal with the current risk spike, or that is what to do now to, to, to, to deal with the, attack a root loop being so compressed, adjust risk calculations and reorient security program resources for increasing volume of patches, decreasing time to patch, and more persistent complex attacks. Nobody thought of this before, I guess, right?

**Chris Romeo (12:51)**: It's news.

**Izar Tarandach (12:51)**: It's okay. What do I do now? It's Monday morning. What do I do now?

**Matt Coles (12:56)**: You panic first, ask questions later.

**Chris Romeo (13:00)**: Where's my towel?

**Izar Tarandach (13:02)**: Where's my towel?

**Matt Coles (13:04)**: Forget that keep calm and patch on thing. You gotta panic first.

**Izar Tarandach (13:11)**: Focus on the basics and harden your environment further. Here I'm going to open a parenthesis. Lately I've been hearing this a lot from my boss, Uri Friedman, with the security brutalism thing that I already told you guys about. Going back to the basics is the only thing that you can do, right? I mean, reduce your attack surface and then you don't have to patch anything.

**Matt Coles (13:33)**: So how far back? So let's be careful about— let's be cognizant of this. How far back? I'm thinking, are we back into the perimeter? There's a perimeter and network segmentation as opposed to everything is open all the time.

**Izar Tarandach (13:47)**: Well, get the perimeter reduction, put it together, zero trust, and every single thing becomes its own perimeter.

**Matt Coles (13:54)**: That's the theory, but that's why you have to patch everything because everything is its own premise. As opposed to network segmentation where you literally wall off parts of your network.

**Izar Tarandach (14:03)**: Exactly. But that's the thing. Network segmentation also depends on the implementation of a TCP/IP stack, which could have its own bugs. So you could be trusting a segmentation that actually doesn't exist. But the moment you start looking at it as a simplification of everything, And instead of just patching everything, you start taking a more critical look at saying, do I really need this thing?

**Matt Coles (14:29)**: Protection first.

**Chris Romeo (14:29)**: Everything.

**Izar Tarandach (14:30)**: Yeah. Is everything that's in my attack surface, is it just fluff or does it have a practical purpose? And that's something that we don't like to do.

**Matt Coles (14:41)**: But that's what we do during threat modeling to some extent.

**Izar Tarandach (14:44)**: That's what we do during threat modeling because threat modeling is the answer to everything. Right? So I could call threat modeling 4G too, and it would sound the same. That's 2 Hitchhiker's Guide to the Galaxy references in less than 5 minutes. I want a bonus.

**Matt Coles (15:03)**: Sure, we'll give you that $0 bonus on top of the $0 we pay you.

**Izar Tarandach (15:06)**: Thank you. So my point, and actually Uri's point. Is that before you do all the, the acting panic runaround, just reduce the frigging, the frigging attack surface. Chances are you don't need everything that you need. Look at the plumbing, make the plumbing simpler before, before you go make it shinier.

**Chris Romeo (15:30)**: Right.

**Matt Coles (15:31)**: And before you think of, before, and, and just to extend that, I guess a little bit, if you think that that thing is, that thing, that one of those targets is useless, it's always useful to an attacker. 'Cause exactly because pivoting and, and, and, uh, you know, APTs exist.

**Chris Romeo (15:46)**: And that's one of the things this paper is highlighting is that, uh, Mythos is, is chaining together vulnerabilities, right? So it's not just finding critical zero-day style exploits. It's doing a lot of vulnerability chaining to get these results. Right.

**Matt Coles (16:03)**: And so those relatively low severity, relatively low severity vulnerabilities that had very little You know, impact as a result, you know, malware authors aren't using them to necessarily, you know, do ransomware attacks. Mythos is able to chain those together to gain a foothold or to gain little bits of information about memory or whatever. Right.

**Izar Tarandach (16:26)**: So basically what you guys are pointing at right now in my head is that we used to, to use things like CVSS to say that something was a medium problem. Because the complexity was high. And MITRE can just be able to reduce that complexity so that what was medium yesterday becomes high and critical today.

**Matt Coles (16:46)**: There's a, there's a metric in CPSS4 though, which is not just the complexity of the exploit, which would be something that Mythos can reduce, but, but conditions on the attack, which Mythos can't control.

**Izar Tarandach (17:01)**: True, but Those are mostly fixed because a given piece of software has a given way of being installed.

**Matt Coles (17:09)**: Agreed. And it, but it's variable by install. In other words, it's, it's fixed based on, uh, there's, there's a, there's a cap to that because of the way the software or, or system, not just software, right? May be utilized or, or may have been, you know, developed and recommended to be installed. But then as each sys, as each company and their organization, then they create the specific deployment that may be variable across deployments.

**Chris Romeo (17:34)**: So does CVSS even matter in this new world of mythos-generated vulnerabilities?

**Matt Coles (17:41)**: That's an interesting question.

**Izar Tarandach (17:42)**: I think it matters. That's an interesting question. In my opinion, it just means that the scale now is going to slide, but I'm very curious to see where Matt was going.

**Matt Coles (17:54)**: So, uh, I guess on the one hand, it's still, I think, valuable in terms of understanding how bad a vulnerability could be. Um, because there's more metric, there's more, more than just simply exploitability in that metric. Uh, so there is, there is that, I guess, but if everything is, if everything's critical, nothing is critical.

**Izar Tarandach (18:18)**: Yes.

**Matt Coles (18:19)**: But Mythos doesn't make a vulnerability critical. Mythos makes a vulnerability easy. Uh, and so it doesn't rate, it doesn't change its severity. It reduces some of the factors and it reduces some of the factors involved in determining that severity, at least under CVSS. Um, but it does, it doesn't change, it doesn't change everything about the, about the vulnerability. So I think the vulnerability calculation is still useful. I might ask an alternative question. What's the value of EPSS? Exploitability Prediction Score.

**Chris Romeo (18:54)**: Hmm.

**Izar Tarandach (18:55)**: True.

**Matt Coles (18:56)**: If everything is exploitable, it's always 1.

**Chris Romeo (18:59)**: Yeah. So it's not, so it just doesn't matter anymore if AI can determine. Go ahead.

**Izar Tarandach (19:07)**: One, one observation on top of that is that we are assuming that everybody has access to something like Mitos, right? Which is probably not going to be true.

**Matt Coles (19:18)**: Well, it's interesting. I'm looking at the table on page, uh, starting on page 17, a Mythos-Ready Security Program Risk Register. And number 1 is accelerated threat exploitation. AI models have discovered, have been discovering vulnerabilities for over a year. Mythos accelerates it, but the capability predates it.

**Izar Tarandach (19:36)**: Yes.

**Matt Coles (19:36)**: So that means if it's not Mythos, it's something else.

**Chris Romeo (19:40)**: Yes, but, and it's going to get rolled into the, it's going to get rolled into the core models. It's just a matter of time.

**Izar Tarandach (19:47)**: I, I don't know. The, the moment that this amount of, uh, red lights start showing up, people go up and say, okay, this is a nuclear weapon level thing. We need to control it. And sure, there will be escapes. There will be other people reproducing parts of it and this and that and the other one. And there'll be some kid buying plutonium on eBay and building their own reactor and whatever. But at the end of the day, if somebody decides that, uh, they need a very good reason to regulate access to something like this, this is it. This is basically a bunch of people very, very, very high in the industry saying, uh, we need controls over this because otherwise it's going to be hell.

**Matt Coles (20:26)**: Yeah.

**Chris Romeo (20:26)**: But we have, they never actually said that in the paper though. We have governments putting words in their mouth.

**Matt Coles (20:30)**: We have governments which are actively funding this type of research, right? So it's, it's not gonna, it's, this is a commercial enterprise, right? Mythos is a commercial company. They have profit and loss margins to worry about.

**Chris Romeo (20:40)**: Well, Anthropic is a commercial company.

**Matt Coles (20:42)**: They just got blacklisted by the government, by the US government, right?

**Izar Tarandach (20:45)**: But to, to, to Chris's point, yes, I'm putting, uh, uh, words in their mouths. That's true. But what I'm interpreting here is a lot of, uh, uh, people that we have learned to really listen to because they have had repeated successes in the things that they have predicted are now predicting something that we as an industry have always put forward as one of the big blockers, which is the time to patch. And they're saying, hey, the only way that you can solve this thing is by raising this, this other thing that you always said that it's very difficult.

**Chris Romeo (21:20)**: Yeah.

**Matt Coles (21:20)**: Now it does say—

**Izar Tarandach (21:21)**: in other words, there's no— in other words, there is no solution.

**Matt Coles (21:24)**: Yeah. It does say non-frontier open weight models can already achieve much of this at an accessible cost. Right. Frontier models are acceleration, not the starting gun.

**Izar Tarandach (21:38)**: Exactly.

**Matt Coles (21:38)**: And each patch also becomes an exploit blueprint. So this is something we've known for a long time, that when patches come out, they get reverse engineered and the attackers know what's being patched. Because you can't hide it, right? I mean, you really can't, you can't really obfuscate a patch well enough unless, unless you encrypt every patch. And it's perfect encryption. And only the systems that can apply the patch can decrypt it. If, a bunch of ifs, right? So patches contain exploits. Patches contain information that exploit writers can use. And we know that LLMs are great at decompiling.

**Izar Tarandach (22:24)**: Now, another interesting thing in here for me, it's, I just got to the table with the priority actions. The priority actions, the defensive ones themselves, are extremely dependent on LLMs, right? They're saying, turn agents and LLM capabilities inward on your own code and dependencies. That there is a big assumption in here that the solutions that we have had up to now, for example, reading code, are just not enough anymore. You need to have an LLM-based one.

**Matt Coles (22:55)**: Well, that's number 2, require AI agent adoption.

**Izar Tarandach (22:59)**: But, but, but that's on the writing. That's on the writing new code. The first one that they say is take an LLM-based scanner, turn it into your code right now. Basically, in, in, in my head, what it means is you can only fight fire with fire. So if the problem is an attack, you have to have a defense, defensive LLM.

**Chris Romeo (23:21)**: If, if, if, why, why is AI generating 500, uh, let's just assume that it is in, in fact able to generate 500 critical, proven, remotely exploitable vulnerabilities against your product. How, I mean, it would take humans a year to fix that, to patch that.

**Izar Tarandach (23:41)**: So their solution is turn your LLM scanners now into your code. A, as if people haven't been doing this already. B, as if we don't have the non-LLM scanners working already. And C, as if magically right now these scanners that didn't find something before are going to start finding things just because Mythos. They are still the same models. They are still the same things.

**Matt Coles (24:05)**: It's calling out non-Mythos, uh, uh, LLMs here, agents, right?

**Chris Romeo (24:09)**: Right.

**Matt Coles (24:10)**: So, and including some open source ones, which is, which is good actually to see.

**Izar Tarandach (24:13)**: Correct. It, it, it's awesome. It's awesome. Right. But to say that this is a critical thing to do this week, because otherwise, huh? I, I, I, I, I'm, I'm, I'm failing to see here how the, the proposed solution is any better than just keep doing what you should already be doing anyway.

**Chris Romeo (24:35)**: Okay. So you're saying your argument is that people should be, are, have already been doing this before this paper was ever written, or they already are doing it in reality.

**Izar Tarandach (24:43)**: No, my argument is, I think, two-headed. A, that this is not something that people should start doing because they should already be doing it. But B, more important, just because a more capable model came up with a bunch of findings doesn't mean that less capable models that existed before all of a sudden have more weight to what they could find or not. So if you, in other words, if you were doing this thing before, using the same models, you're not going to start automagically finding things this week because Mithras found things last week. In other words, this is just something that sounds good, but it's not immediately going to give you any, any added value.

**Matt Coles (25:34)**: If you weren't, if you weren't already doing it, it gives you a ton of value. If you were already doing it, you're not gaining anything by doing it. You're not going to gain anything by doing it more.

**Izar Tarandach (25:44)**: Right.

**Matt Coles (25:44)**: By continuing to do it, I should say. Right.

**Izar Tarandach (25:46)**: And if you were not doing it before with either an LLM-based or non-LLM-based scanner, then seriously, what are you doing?

**Matt Coles (25:54)**: I mean, the theory, I mean, the theory goes here that the LLM scanner is better than the commercial SaaS tool.

**Izar Tarandach (26:03)**: Have we seen numbers saying that they are?

**Matt Coles (26:05)**: I don't know.

**Izar Tarandach (26:06)**: I'm genuinely asking.

**Matt Coles (26:09)**: I don't know. It's a good question.

**Chris Romeo (26:10)**: Yeah. I don't know either.

**Matt Coles (26:12)**: I mean, if you had a, if you had a, like if you had a Coverity or Fortify from ages back, right? I mean, that those were, those were high performance, you know, uh, very noisy, but high performance static analysis tools, right? They looked at call stacks and, you know, variable flow and the whole works, right? Is LLMs better? They're reading the code directly, whereas those SAS tools were looking at compiler output. Now, granted, we're talking about compiled languages versus interpreted languages. Maybe the language pipelines are better or worse. I think the SAS tool capabilities have changed over time. And of course, now we also have, you know, a plethora of supply chain attacks or, you know, dependency problems, which maybe LLMs are better at making those connections than the tools themselves.

**Izar Tarandach (27:15)**: That's the thing, the making the connections. You touched exactly at it. I would say that nowadays, if you're not using both approaches at the same time, And the specific reason why I say both is that, first of all, before the whole LLM thing came true, we already knew that we were having very, very good non-LLM scanners in terms of building all the AST, understanding the language, understanding how things are written and not, understanding the artifacts that even interpreters and compilers could put into the final code that gets deployed. But the most important thing that they give us is determinism. They do go through all the paths. They do go all the good ones, at least.

**Matt Coles (27:58)**: Yeah.

**Izar Tarandach (27:58)**: They visit all the, the branches in the tree.

**Matt Coles (28:01)**: When given enough resources, they do.

**Chris Romeo (28:03)**: Exactly.

**Izar Tarandach (28:03)**: Yeah, exactly. It's resources and time. But, well, time's a resource. But, uh, uh, once you have that shopping list, giving it to an LLM to check one by one, that's a different use case than just typing on your agent. "Find some vulnerabilities in my code." Is it that simple?

**Matt Coles (28:22)**: Really? Can I do that right now?

**Izar Tarandach (28:24)**: A, you can. B, yes, some people do that. C, no, that's not the right way of doing it. But the important thing here is that once those things come together and become one pipeline, they are much stronger than working apart, right? So again, we go back to the AI-infused thing. Not AI-powered thing. You have to have AI helping your other tools, helping your other methodologies, and not always, in some cases, yes, but not always substituting them altogether. So Captain Obvious, la la la la la here, but things evolve. They don't always, uh, uh, surpass each other and, and don't always take one input, take one output, one in, is the best move here.

**Matt Coles (29:18)**: Okay, so I'm looking at this list of priority actions. And something, something is troubling. Point your agents at your code. Number one, require AI agent use across your, your, across your organization. Establish governance and defend your agents. So go ahead and deploy this technology that has, that which they say specifically has poor defenses in effect, right? Without agents, most tasks will be untenable, but they must be defended. Agents are not covered by existing controls and introduce both cyber defense and agentic supply risk. So make these things ubiquitous in your organization. Prepare to be attacked.

**Chris Romeo (30:17)**: Yeah.

**Matt Coles (30:18)**: Not, listen, not build your infrastructure and defenses first, then deploy.

**Chris Romeo (30:23)**: Well, this is the AI age, Matt. Nobody has time for that.

**Matt Coles (30:28)**: Boy, it's like, it's not like we have to patch or anything.

**Izar Tarandach (30:35)**: Yeah. At the end of the day, I, I think that nobody is going to go into a, uh, what's the name of those processes? The, um, digitalization, uh, uh, jump or whatever. Whenever, whenever everybody goes to the cloud and those massive lifts and all that.

**Matt Coles (30:51)**: Oh, digital transformation.

**Izar Tarandach (30:52)**: Nobody's going to, yeah, no, nobody's going to go like the other way around and say, um, oh no.

**Chris Romeo (30:57)**: I'm still holding off on the cloud. Don't trust it.

**Matt Coles (31:01)**: I heard rumors you might be better secured than everyone else.

**Chris Romeo (31:06)**: Uh, so just to close the loop on this list, like, I mean, I think this list of priority actions is fine. I mean, nobody ever takes a document like this and says, well, I better get to work. Team, you'd be surprised. Point agents at your code and pipelines.

**Matt Coles (31:22)**: Nope.

**Chris Romeo (31:22)**: And then we have to require AI. Like, people don't—

**Matt Coles (31:24)**: this is an action list for CISOs. That this is, that's what, because it's because of the names and because of Mythos and people are just gonna do this blindly. I, I have, I'm, I will predict that many companies will implement these almost verbatim.

**Izar Tarandach (31:45)**: And I'm going to predict on top of Matt's prediction, two things. First of all, next week we're going to see some kind of standard that verbatim goes through the steps and says, this is the new reality.

**Matt Coles (31:57)**: A, NIST 800-53 Rev 6, the LLM version.

**Chris Romeo (32:02)**: No, it would take them at least 5 years to write that. Come on.

**Izar Tarandach (32:06)**: And, and, and B, I want to say, I, I wanted to, to, to go back to something that Matt said before the microphone went on. If anybody wants to try to implement this as is in this table, Cybersecurity industry as a whole is saved, right? Because they will need so much, so much, so many, so many people. Thanks. Cause it's countable now. Yeah. So many people to implement this thing that everybody, anybody that's sitting down in front of a computer nowadays and saying AI is going to take my job, those people are going to have new jobs real, real, real fast.

**Chris Romeo (32:49)**: Framework installers.

**Izar Tarandach (32:51)**: Yeah, you're going to need so much work.

**Matt Coles (32:56)**: You need humans to make the decision. So number 7 is our earlier point about inventory and inventory and reduce attack surface.

**Izar Tarandach (33:06)**: Yep.

**Matt Coles (33:06)**: In order to do reduction, you have to make a decision. And unless you, well, I guess until the agents are good at automatic targeting of systems, then, and understanding value of those targets, then you need a human in the loop. And you need a human, you need a human or humans in the loop who can understand the environments and what the attack surface represents and the value of those systems as well as the value to the attacker. The AI can assist, but the human, I think, will probably be necessary.

**Chris Romeo (33:42)**: At least in the short term.

**Izar Tarandach (33:43)**: Because if your AI can do all this stuff by itself, you have different problems.

**Matt Coles (33:48)**: But, well, yeah, I mean, a big, a big, uh, power bill, probably, and token bill.

**Izar Tarandach (33:54)**: That too. But it's fun that Matt pointed at the, at the item 7, and I went to the big text in there and immediately something like jumped up.

**Matt Coles (34:04)**: As fun.

**Izar Tarandach (34:04)**: Right in the middle of the page.

**Chris Romeo (34:06)**: As fun.

**Matt Coles (34:06)**: As fun jump. That's where my eyes went to.

**Izar Tarandach (34:09)**: And I was going, wait, wait, wait, S-bombs. And I read the whole sentence, generate real S-bomb.

**Matt Coles (34:15)**: This is on real S-bombs.

**Izar Tarandach (34:17)**: No, no, no, no, no. And then I went to next sentence because I was hoping that it was going to tell me what to do with the real S-bomb that I just generated. It doesn't.

**Matt Coles (34:26)**: No. In fact, it's back, fast backwards. You should be aggressively shutting down things, then generate your S-bomb.

**Izar Tarandach (34:31)**: And then do the S-bomb.

**Matt Coles (34:33)**: But, you know, it is, it is open for public comment. We should. Submit that comment.

**Chris Romeo (34:37)**: Just put it out there. Okay, so here's the ultimate scary question.

**Matt Coles (34:43)**: Oh, isolate. Sorry, isolate or air gap at risk systems. It is talking about network defense.

**Chris Romeo (34:49)**: We're going to air gap. We're going to air gap. In the modern internet, we're going to air gap.

**Matt Coles (34:52)**: We're going to go for enclaves again. Enclaves are coming back, baby.

**Izar Tarandach (34:55)**: Oh, remember the enclave?

**Matt Coles (34:57)**: Remember the enclaves?

**Izar Tarandach (34:58)**: Remember how the enclave got popped?

**Matt Coles (35:00)**: Remember all the Tempest research and how you could break enclaves? Yeah.

**Chris Romeo (35:05)**: True. That's getting really nerdy though.

**Matt Coles (35:07)**: Yes, it is.

**Chris Romeo (35:08)**: When you, we start getting into that level. So here's the question for you, philosophical, so that you can philosophize, to use a word that we shared earlier.

**Matt Coles (35:17)**: If you can spell it, that'd be awesome. But I can't, I can't, I can't say, but I can't say mythos.

**Chris Romeo (35:23)**: It's mythos. Mythos.

**Matt Coles (35:24)**: Mythos.

**Chris Romeo (35:26)**: Mythos. Mythos. Can't, do we reach a world where all the critical vulnerabilities are found for a product and mitigated. So there are no longer— so you have a product and it no longer has critical vulns.

**Matt Coles (35:43)**: No. It depends why some vulnerabilities are unpatchable.

**Izar Tarandach (35:52)**: Yes.

**Matt Coles (35:53)**: At some point you will run out of the technical vulnerabilities, right? You'll run out of the buffer overflows.

**Izar Tarandach (35:59)**: Okay.

**Matt Coles (35:59)**: And you'll still be left with business logic flaws that you probably cannot patch.

**Chris Romeo (36:06)**: Why wouldn't you be able to patch a business logic flaw?

**Matt Coles (36:08)**: Because it's human-generated, human-informed, uh, in other words, in order to make the system useful for a target audience, there are almost certainly going to be flaws that you cannot patch.

**Izar Tarandach (36:22)**: True.

**Matt Coles (36:23)**: Authentication issues, authorization issues. Privacy problems. You're not going to necessarily be able to patch all of these.

**Izar Tarandach (36:31)**: On top of that, you're assuming that every patch works 100% and doesn't bring problems with itself. And we have seen again and again and again that the patches work.

**Chris Romeo (36:39)**: That's why you have AI quality assurance.

**Matt Coles (36:44)**: If that patch breaks major functionality, like, hey, here's how I patch this. I turn off this API. You're not gonna do that.

**Izar Tarandach (36:55)**: No, but, but that's saying that the only way to patch is to remove functionality, right? So it's taking the worst case that you, you pointed out, right? Yeah.

**Matt Coles (37:03)**: At some point you'll run into patches that require significant modification to the, you know, of the system.

**Chris Romeo (37:10)**: At the end, there's the way AI writes code. It, it operates a whole lot faster than human beings do, so it can just rewrite Uh-huh.

**Matt Coles (37:19)**: And if one system— And open new attack surfaces as a result.

**Izar Tarandach (37:23)**: But you're assuming that every system lives in a vacuum. Systems talk to systems. If both systems don't go up at the same time, then one system could hold up the patch of the first one and say, well, if I want this functionality, I can't patch because otherwise I can't talk to the other system.

**Matt Coles (37:42)**: This is the problem of containers, right? When you spin up and down containers, If one comes up on a different rev, you've broken your infrastructure.

**Izar Tarandach (37:51)**: It's the whole dependencies, not dependability, dependencies thing, right? Things have to move forward at the same pace that they maintain compatibility with each other.

**Matt Coles (38:03)**: And the big trend right now is pinning your dependencies. So even if you apply a patch, you're still not going to get the patch because, I mean—

**Chris Romeo (38:11)**: That was famous 10 years ago. Dependency padding.

**Matt Coles (38:15)**: Okay. It still continues today.

**Izar Tarandach (38:18)**: Yeah. Yes.

**Matt Coles (38:21)**: Uh, so number 9 on this list, I think, is, uh, will require some discussion and we, I don't think we have time for it today, but at some point we probably should talk about it.

**Izar Tarandach (38:31)**: Wait, which one? 9?

**Matt Coles (38:32)**: 9.

**Chris Romeo (38:33)**: Build us a deception capability.

**Izar Tarandach (38:35)**: Yeah. So bring on the Decepticons.

**Matt Coles (38:42)**: So deploy canaries, honey tokens. I don't even— and actually, if you guys know what honey tokens are, let me know.

**Chris Romeo (38:51)**: I do. I do.

**Matt Coles (38:53)**: What's a honey token?

**Chris Romeo (38:54)**: Unfortunately, it's like a fake password. Oh, or, or token that gets left somewhere so that somebody can find it and then use it, and then the use of it triggers the detection.

**Izar Tarandach (39:06)**: So it basically becomes a canary. So I don't see why the distinction between both things, but the truth is canaries, especially at the, the secret and credential level are really useful. My question again comes to time. Do I deploy these things before I do the other, the other 10, 11?

**Matt Coles (39:26)**: No, it says, yeah, the other two things, it's high. So do it in the next 90 days and 6 months out.

**Izar Tarandach (39:32)**: Yeah, but 9/11 is critical in the next 6 months. So how does it work? It's critical, but do it in the next 6 months.

**Chris Romeo (39:40)**: They should have sorted this by start.

**Matt Coles (39:42)**: But also, what's— what is VulnOps?

**Izar Tarandach (39:47)**: It's a new ops that's going to generate a completely new industry.

**Chris Romeo (39:51)**: Watch this space. I've heard that term thrown around for a year or two at this point. VulnOps.

**Matt Coles (39:56)**: Autonomous Vulnerability Research and Remediation. Isn't that Pen testing?

**Chris Romeo (40:02)**: No, because it's not— pen testing is not autonomous at this point.

**Matt Coles (40:05)**: Not yet.

**Chris Romeo (40:08)**: That's the, that's the car crash that we're heading towards.

**Izar Tarandach (40:11)**: Staffed and automated like DevOps, but for autonomous vulnerability research and remediation.

**Matt Coles (40:16)**: So this is a team within your development staff that just finds vulnerabilities?

**Chris Romeo (40:21)**: That used to be QA, you know, but it's using Mythos to find all those vulnerabilities before external attackers can find them.

**Matt Coles (40:30)**: Well, I thought your agents were doing that before when you started the program.

**Chris Romeo (40:34)**: Someone's got to run the agents. Agents aren't completely independent yet.

**Izar Tarandach (40:38)**: So you Google for VulnOps. The first thing that comes up is a project on GitHub. Then the second one is a domain saying VulnOps coming soon.

**Chris Romeo (40:49)**: Then we have standby for VulnOps.

**Matt Coles (40:53)**: Uh, go for VulnOps.

**Izar Tarandach (40:55)**: Oh, the third one is actually a LinkedIn post that says what it is. Let's see what it is.

**Chris Romeo (40:59)**: We could rename our, our podcast, the VulnOps Podcast.

**Matt Coles (41:03)**: Well, then there'd be no conflict and we'd be the original VulnOps Podcast.

**Chris Romeo (41:06)**: Well, then somebody would just copy us. Somebody would just make another podcast called the VulnOps Podcast. You know how it goes.

**Izar Tarandach (41:14)**: Much like how MDR brings together a plethora, plethora of technologies to help defensive teams detect and respond to threats, VulnOps does a similar job for vulnerabilities.

**Matt Coles (41:29)**: Okay.

**Chris Romeo (41:30)**: I'm liking it so far. I'm gonna get a t-shirt.

**Izar Tarandach (41:32)**: Yeah, but that's the thing. I don't care about being corrected. I'm used to it.

**Chris Romeo (41:37)**: No, no, no. I'm just saying, I'm talking about this whole idea of Olmops. I think I might be in.

**Izar Tarandach (41:41)**: Oh, yeah, yeah, yeah. No, it's just that the thing promised to tell me what it was and it didn't. So I'm still looking.

**Chris Romeo (41:48)**: Typical AI.

**Izar Tarandach (41:49)**: No, no, this was normal Google.

**Chris Romeo (41:52)**: Typical humans. Humans. Vulnerability.

**Matt Coles (41:57)**: So, so at the end of this, is this something to watch or something to take action on?

**Izar Tarandach (42:03)**: Which one?

**Matt Coles (42:04)**: This paper thing, the whole thing.

**Chris Romeo (42:06)**: No, no, no, no.

**Izar Tarandach (42:07)**: Definitely something to watch.

**Chris Romeo (42:08)**: I think, I think you gotta take action, but I think you gotta start rocking and rolling right now if you're, if you want to get left behind.

**Matt Coles (42:17)**: Yeah.

**Chris Romeo (42:17)**: I mean, you can't, you can't roll this type of stuff out in a weekend. This is not a weekend project.

**Matt Coles (42:22)**: Wait, you can't firecode your security programs?

**Chris Romeo (42:25)**: I mean, you should. Somebody should try it so we could laugh at them.

**Izar Tarandach (42:27)**: Not again. But, uh, no, no, no, no. I think that, I think that first of all, in my opinion and following Security Brutalism, you have to go and ask yourself, why haven't I been doing half of this already? Right? Then it's the reduce, reduce, reduce, reduce. Seriously, just reduce. I look around my room in here and the amount of old digital junk that I have at some point I should throw away, but sentimentality hits in and I can't, but, uh, don't be sentimental to your system.

**Chris Romeo (43:00)**: Certainly the attack surface reduction is not a bad thing, but it doesn't solve this problem.

**Izar Tarandach (43:05)**: It doesn't solve, but in the amount of things that you need to actually fix and keep an eye on, it brings you back to fundamentals.

**Chris Romeo (43:12)**: Right. I mean, I've been preaching attack surface reduction for at least a decade at this point. And the funny thing is nobody ever does it.

**Matt Coles (43:18)**: No, because the more computing resources you find, you get, the more you use.

**Izar Tarandach (43:24)**: Because budget, the more you get to buy the next shiny thing, you're going to get more budget next year. Right? So it, it, it's apparently it's better to say I need $100 million to get the next shiny thing than I saved $100 million by dropping that shiny thing from last year. Mm-hmm.

**Chris Romeo (43:43)**: Yeah.

**Izar Tarandach (43:44)**: Right.

**Matt Coles (43:44)**: Which means, which means on this list should have been budgets should change. Budget planning should change.

**Izar Tarandach (43:50)**: Yep, definitely.

**Chris Romeo (43:51)**: They never talked about budget, did they?

**Matt Coles (43:54)**: Nope. They didn't say how much this is gonna cost. Right. We don't know what the token cost for Mythos is and how, what the myth, what the token budget is.

**Izar Tarandach (44:04)**: And how many nuclear sites we're going to have to build to power this whole thing.

**Chris Romeo (44:08)**: Hopefully a couple.

**Matt Coles (44:10)**: I'd rather build a nuclear site at this point than go back to coal mining.

**Izar Tarandach (44:18)**: True. So what do we close with this thing? Talking point: an aggressive plan is needed.

**Chris Romeo (44:25)**: Um, I could say that about anything though.

**Izar Tarandach (44:28)**: No, wait, wait, wait.

**Chris Romeo (44:29)**: They say aggressive plan is needed about mowing my lawn. I need an aggressive plan.

**Izar Tarandach (44:34)**: Wow. The funded foundation is why our program can adapt rather than react in a crisis. What has changed is—

**Chris Romeo (44:45)**: yep, I was making up my own OODA on the fly.

**Matt Coles (44:51)**: Defender agents, do not wait for industry governance frameworks. Defend— define your own now, because that always worked.

**Chris Romeo (45:04)**: Well, so I guess we leave this as a wait and see.

**Izar Tarandach (45:07)**: Actually, they have a closing paragraph here that I really like. Building a MITRE-ready security program is not about reacting to one model or announcement. It is about permanently closing the gap between how fast vulnerabilities are found and how fast your organization can respond. In other words, reduce your OODA.

**Chris Romeo (45:25)**: Does that actually say reduce your OODA, or are you just adding that in?

**Izar Tarandach (45:28)**: No, no, no, that's me. That's me.

**Chris Romeo (45:30)**: Okay. I thought that was what I was going to hear, but I was just checking.

**Matt Coles (45:33)**: That's me.

**Izar Tarandach (45:33)**: Yep. That's me. That's me.

**Chris Romeo (45:36)**: And so, I mean, it's not, they're not, they're not saying anything we really disagree with at the end of the day. Nope.

**Izar Tarandach (45:41)**: Nope.

**Matt Coles (45:41)**: This is double down and make it happen.

**Izar Tarandach (45:43)**: I mean, yep. Bordering on duh stuff.

**Matt Coles (45:47)**: And, and, and this is also, I guess, I guess I've come around on this paper. It, I was a little skeptical reading it to start, but it's really, if you are resistant to any of this, figure out a way to not be resistant to this.

**Izar Tarandach (46:03)**: Yes.

**Chris Romeo (46:03)**: Because as the Borg are famous for saying, resistance is futile.

**Izar Tarandach (46:10)**: Futile.

**Matt Coles (46:11)**: You will be assimilated.

**Chris Romeo (46:13)**: You will be assimilated. I think that's where we leave it.

**Matt Coles (46:16)**: I think that's Star Trek is much better at for these than Star Wars in my opinion.

**Chris Romeo (46:24)**: But live long and prosper. That's right. That's— ISAR is the OG Spock threat modeling supermodel.

**Izar Tarandach (46:33)**: Oh yeah.

**Matt Coles (46:36)**: Wait, wait. You did.

**Izar Tarandach (46:39)**: Hey, it comes from the same place, right?

**Chris Romeo (46:43)**: Right from the heart. It's from the heart.

**Izar Tarandach (46:45)**: I got membership. I got membership on the tribe. Yeah.

**Chris Romeo (46:48)**: All right. Well, folks, we'll leave this with you here. Go basically build the whole ALI Vaultnops program over the weekend, and we But reduce your, reduce your attack surface before. Reduce the attack surface before. And we expect a TPS report diagramming the whole thing.
