What Is Mythos? The AI That Found a 27-Year-Old Vulnerability
Mythos is Anthropic's experimental AI-powered vulnerability research system. In testing, it identified a 27-year-old OpenBSD vulnerability and developed a working exploit with minimal human involvement
It gained attention after Anthropic reported that Mythos uncovered a 27-year-old vulnerability in OpenBSD and generated a working exploit during testing. More broadly, Mythos has been presented as evidence that advanced AI systems can perform vulnerability research at a scale and speed that will significantly change both defensive and offensive cybersecurity.
In this episode of Open Source Open Mic, we break down what this means, how it works, and why it matters far beyond a single bug. As AI gets better at finding vulnerabilities, the window between discovery and exploitation keeps shrinking. Security teams aren't just dealing with more code anymore. They're dealing with machine-speed software risk.
If AI can find vulnerabilities faster, how do developers keep shipping with confidence?
Subscribe to stay up-to-date on the latest.
Transcript
0:03: It can find software vulnerabilities that have survived years of human review, and it's so powerful that access is being tightly controlled.
0:12: Today we're unpacking Mythos, the AI model that's forcing the cybersecurity industry to rethink what's possible.
0:19: Let's break it all down on this episode of Open Source, Open Mic.
0:26: This is Open Source Open Mic, the podcast where we talk about everything happening in the world of open source security.
0:32: So pull up a chair, it's time for an open conversation.
0:37: Hello everyone and welcome to another episode of Open Source Open Mic.
0:41: My name is Andrew Garrett, and I work in product marketing here at Sonotype.
0:46: I'm pleased to be joined today by Ilka Turinen, and Ilka is the field CTO here at Sonotype.
0:55: He's based in the UK in London, and so I'm, I'm glad to have him for a few minutes.
1:00: It's the afternoon for him, it's the morning for me here in Utah, but we're gonna have a great conversation today.
1:06: Ilka, is there anything else you'd like to say about yourself, about your background?
1:11: really, just that I'm really excited to be here.
1:13: I'm really looking forward to the conversation.
1:14: I've obviously been with, Sunside for 11 years.
1:19: I've been working quite a lot, seeing the sort of software supply chain.
1:23: , terminology and phenomenon that really come together over time, so we kind of started from Debsecops and moved on into sort of serious at scale supply chain management as an industry throughout the course of these years.
1:38: So, so I'm really have been, sort of working with customers and organizations and governments.
1:45: on, on, on it along the way, so it's a, it's a pretty interesting topic, what we're gonna be talking about today, so very excited.
1:55: Awesome, well, it's great to have you here, Ilka, and our topic today is gonna be all around mythos.
2:02: we, we've been hearing a lot about it, mythos is, Basically, Claude's version of a, of an AI security tool, but I, I wanna hear it from you because you're, you're the expert here.
2:15: Can you give us a quick background on Mythos, what it is and, and why we keep hearing about it?
2:20: Yeah, for sure, for sure.
2:21: I mean, let's start with the very, very basics.
2:25: Mythos, which is by the way, a Greek derived word, which is why it's pronounced mythos, I was told recently.
2:33: is the next version of the anthropic cloud, foundation model.
2:39: So it's essentially, you know, think about Cloud Opus 4.6, 4.7, which are sort of the currently available foundation models.
2:48: This is gonna be the next next spearhead model that's gonna come out there.
2:53: They've been testing it, testing it internally for a while, and they released a, a private beta or private preview preview of it to select it.
3:04: Organizations over the line and in that course of testing what they've realized is that the mythos model continues both on the linear growth of these foundation models to be fairly good at code production.
3:21: so, each of the foundation models, especially over the last sort of 9 months, have really made leaps and bounds in their ability to produce code that starts to be a lot, higher quality than just, you know, sort of, a code assistant type of work, and in the course of over the course of that testing, they also realized that it's a particularly good model at reasoning.
3:44: Especially when it comes to cybersecurity, so there are some, write-ups that they've released, that describe situations where, for example, the model supposedly was completely offline.
3:53: It was left to work on a puzzle.
3:56: they figured out the puzzle and then it let its owner know through an email message saying, hey, by the way, I've solved it.
4:02: Which is a bit of an interesting situation because the model was supposed to be completely air gapped.
4:07: So it had sort of reasoned, it reasoned its way from that air gapped network into an email server somewhere and then, you know, found a way of informing the researcher that was working on it at the time.
4:20: there's been other, organizations playing around with it, you know, testing things.
4:03: So like when that open source gets downloaded and installed on your machine during a build, it'll go reach out and connect to maybe some command center, and then it'll download additional instructions to do certain things.
4:25: Firefox, for example, or the Mozilla Foundation released a blog post where they said that the latest release of Firefox actually contains something like 270, security vulnerability fixes that were discovered using this Mythos model.
4:41: and there are sort of corroborating stories elsewhere.
4:43: In fact, Anthropic says that in our testing, as we've kind of played around with it, it's been able to uncover new and previously unknown security vulnerabilities at a rate that, that, you know, known human researchers have been able to do.
4:56: For example, there's like a famous case of a.
4:59: 27 year old vulnerability discovered in Open BSD that was sort of hiding there, until Mythos was able to sort of piece different pieces of information together and build a working exploit based on that.
5:12: So there's a lot of hype, there's a lot of sort of conversation about, you know, is it this capable?
5:17: I don't think that there's any reason to not believe that it represents a sort of step in function increase in terms of the ability of these foundation models to conduct cybersecurity, you know, activities like vulnerability discovery as well as exploitation, And that of course is, is creating a little bit of tension because that creates an advantage, especially for adversaries because they're, they, if they can discover previously unknown zero days and put them into production faster than ever, that places quite a lot of pressure on organizations to be able to act on those as they are discovered, and do so quickly, so.
6:00: So really I think it's a good example of how the models are evolving, how fast that evolution is now going, and that just means that everything's gonna speed up and you know the level of discovery is gonna speed up the level of code is gonna speed up the ability to produce both attacks and patches for said issues is gonna increase, so.
6:24: We're entering into interesting times and that's why there's sort of a lot of collective anxiety I'd say about, hey, what does this mean, you know, is I feel like we're talking about another Y2K, frankly, it's, it's that sort of like, you know, it's coming, it's, it's happening, and as a result, like there's a lot of focus going on right now on what do we as an industry, as organizations need to do in order to be ready for it.
6:50: Yeah, you know, the hype is real, and, and the anxiety is real, and it's not just in our industry.
6:57: I mean, just the other day I was having a conversation with one of my friends, you know, at, outside of work, he, he's not, he doesn't work in the cybersecurity industry at all, but he knows that I do, and, and he asked me about me.
7:10: And, and what my thoughts were on it, and, I, I would like to pose the same question to you, Ilka.
7:17: how do you think Mythos will shape the cybersecurity industry?
7:20: What, what predictions do you have, about the evolution of where Mythos will take us?
7:26: You know, I think it's a it's a really good question.
7:28: I think I've been asked that question quite a lot over the last month or so, So, like I said, it represents an increase in capability, and what that really is going to mean is that, even if they hold this release back, that doesn't mean that models won't reach that capability elsewhere.
7:48: There's already indications that the latest Codex models from OpenAI as an example, are able to do similar cybersecurity type activities.
7:57: activities at the same level as the mythos model.
8:00: So, so, I think it's a reasonable, reasonable estimate that within the next 6 to 12 months other model families are, are going to catch up in terms of capability, meaning that this is just gonna be the next new watermark for for models, and their ability to both produce code and discover these vulnerabilities and execute these attacks.
8:22: That's fact number 1.
8:24: The fact number 2 is.
8:26: It right now it will give an asymmetric advantage to attackers because if they're able to use these next generation models to discover new previously unknown security vulnerabilities and put them into exploitation, then then that obviously represents a lot of risk for downstream organizations, you know, subjected to these attacks.
8:49: and more so if, there's indications that what, what really is going to happen, is that, you know, the level of vulnerabilities being discovered is just going to increase.
9:00: There's already been, you know, if you look at the last 5 years.
9:03: About a nearly 300% increase in the total amount of security vulnerabilities being discovered on an annual basis compared to the beginning of the decade.
9:13: and that it can be largely attributed to AI tools helping that discovery, like literally just the other day, the maintainer of Linux Linus tool vaults said in an email list that their security list currently is unmanageable because so many people are using these AI tools to discover security vulnerabilities, they're all overlapping.
9:32: So from a defender perspective that's gonna put a lot of pressure.
9:35: You're gonna get a lot of bug reports, you're gonna get a lot of notices of new security vulnerabilities.
9:40: So being able to, able to prioritize quickly, hey, what do we do with these, do they affect us?
9:45: Do we have this software?
9:46: Do we have to apply a patch, develop a patch, all of that sort of stuff?
9:50: Has this been even reported to us before, is gonna become really, really important and part of the reason why the collective anxiety kind of exists is that.
9:59: That ability to when we look at historic incidents of like high severity security vulnerabilities, that muscle just hasn't been there, we looked at something like Loch for shell as an example, 5-6 years ago when when that happened.
10:17: you know, 10 out of 10 security vulnerability in an extremely popular, popular open source library, it took the industry about a year to drop below, 40%, download rates, of, of vulnerable versions of, of Lockfo.
10:32: It took a really long time.
10:34: We heard of companies that, you know, took, months and months and months to go back to business as usual as a result.
10:40: That's just one security vulnerability.
10:42: Now imagine that there's a vulnerability of that magnitude every week.
10:46: In something really popular that's sort of the the scale and step increase that might happen but of course on the positive side, you know, if you think about it from a from a sort of less doom and gloom what that also means that is that organizations are able to use these models effectively put them into play, develop patches faster, you know, defend against these new vulnerabilities faster, And use that capability also to apply defense in depth in a way that we probably have never been able to do as an industry.
11:19: So, so either way it sort of represents the direction that the entire software economy is going where we're going into an energetic transformation.
11:27: and from a cybersecurity perspective, it just means there's just gonna be more of everything.
11:31: There's gonna be more security vulnerabilities, there's gonna be more disclosures, there's gonna be more patching required both of your own software as well as your vendor software, and thinking about that new scale, I think.
11:42: You know, if you think about that scale as sort of 50x increase in CVs being discovered, I don't think that that would be far off from what we're gonna see.
11:52: Yeah, it, it sounds like everything's just accelerating because, like you said, we have more security vulnerabilities, but we also have faster patches now, faster time to remediation.
12:03: So, let's talk about that, that pace, you know, for, for those of us that, that are in the industry, how do we keep up with this accelerated pace?
12:14: what, what can we do to, to stay ahead of this acceleration that we're seeing?
12:19: Really good question.
12:20: I mean, you know, first, the bad news, just a couple of, just over a month ago, basically the, foundational, the Security vulnerability vulnerability database, the NVD, the National Vulnerability Database that is run by the US National Institute of Standards, announced that they're reducing their scope of processing security vulnerabilities.
12:41: So it used to be that, you know, you'd open a CVE, then that would.
12:44: Go through the NVD process and you know get things like hey what does this affect what what applications are covered by this vulnerability, what is the severity of this vulnerability, all those sort of foundational things that organizations use to prioritize these vulnerabilities.
13:01: Now what they're saying is we are only going because of the sheer volume of things being reported to us we are only going to prioritize stuff that is used by the US federal government, stuff that is on the known exploitable vulnerability list, which is also a US government list, and things that fall under a specific, you know, criticality designations in the US so what that means really is that almost immediately.
13:26: They said we're not going to process 130,000 security vulnerabilities that were waiting in line to be processed because we don't rank them, so from a defense perspective that's really going to mean that the sources of truth are gonna get more fragmented.
13:39: When, when these models discover these vulnerabilities, they're not gonna wait for CVE to be erased, they're gonna go into production right now, so that means that discovering these vulnerabilities is gonna require, require in-depth mining of bug trackers of issue trackers of GitHub accounts, all of those sort of things which is no minor undertaking I know because that's what we do as an organization.
14:02: One of the key things and capabilities that we've been developing over the years is our ability to monitor the software supply chain at scale, look at all of these things and discover security vulnerabilities when they are discovered by the projects, as opposed to waiting for them to appear on some database.
14:18: So the defense model is going to get a lot more fragmented, so you need to find a way of finding a source of truth.
14:26: Like us, that is able to tell you about the supply chain, that is able to sort of do a lot of that deduplication and that prioritization, that's no longer happening at an industry level.
14:38: Mhm
14:39: Yeah, I'm glad you men mentioned some of the capabilities that Sonotype offers, when it comes to continuous monitoring, making sure that your software supply chain is, you know, clear of any malicious vulnerabilities that, that could be coming in, but, let, let's talk more about some of the best practices, you know, whether, whether you're a Sonotype customer or not.
15:06: What are some of the best practices that everyone should be doing in order to, secure their software supply chains, and embrace this, this new age of, Cybersecurity.
15:20: Well, it's interesting that you mentioned that we just literally wrote a white paper about the subject quite recently because, because I think this is sort of a time not to reinvent the wheel, not to forget the lessons of the past.
15:34: So we as an industry have been spending 20 years or so building these Devsakops models and you know, the good news here is that none of this is a new threat landscape.
15:43: There's certainly.
15:45: Certainly, you know, new types of velocity concerns, but if you really think about it fundamentally, if our agents are writing more code, they're using more open source, our software supply chains are gonna get bigger.
15:56: at the same time, attackers are looking at software supply chains, they're trying to find new vulnerabilities.
16:01: That really means that you need to find ways of getting back to the basics, asking yourself, you know, over the next 30 days or so, even some very basic questions like, do we have centralization of our artifacts?
16:12: Do we have central ingestion points where our developers are downloading open source?
16:16: Do we know all of our vendors?
16:18: Can We acquire S-bombs of the software of all of our vendors?
16:22: Can we produce S-bombs of the software that we write ourselves?
16:26: That's been a conversation we've been having as an industry for the last half a decade, literally since Lock for Shell.
16:31: But now is a good time and it's a great catalyst, to really drill back down into, into those sort of fundamental basics because if a security vulnerability happens tomorrow.
16:41: You know, the incident's gonna go like this, hey, which applications are affected?
16:45: Do we know how to patch it?
16:46: Do we know what versions are safe?
16:48: Do we know, know what versions are not safe, And as a result, as a result, even just doing that basic kind of tabletop exercise or even running a fire drill can really, really, uncover some of those sort of fundamental questions like do we even have the capability of this?
17:06: Would we, how would we get alerted about a new security vulnerability tomorrow in one of the open source components that we have?
17:12: Are the tools that we use just leveraging these open databases, are they, are they genuinely monitoring these things live or not?
17:20: That's question number one.
17:22: There's, there's also the point of setting some KPIs of what do you think should happen when, you know, an incident like that happens, you know, if we assume that one's happening every week, then how are we, what is the MDTR that we're gonna be shooting for?
17:38: Is it, is it a day, is it a week, is it a, is it a couple of days?
17:42: I don't think there's a right answer per se.
17:45: But you do need to have a little bit of a think of of what the strategy here is, in this supply chain incident type of you know, type of world, And then basic things like who's responsible for what, what's who, who's responsible for this application if an incident happens, we triage the right application, we know where it's affecting us, who does the fix?
18:06: Do we go in and fix it by ourselves?
18:08: Do we, do we instruct the team who's the responsible party?
18:11: A fire drill is a wonderful thing to really shake those sort of questions out.
18:15: It really kind of gives you a little bit of a sense of sense of ownership and control.
18:20: for sure, but over the next 30 days, that's literally what we're, what we're advocating people to do is really think about those basic control points that you have.
18:29: Think about the basic individuals assigned to these things.
18:32: Think about the prioritization mechanism you'd have if you think about this new scale, and then find ways of proving or not proving that you have this level of coverage.
18:43: Yeah, I think that's a great point, because, like you said, we're not reinventing the wheel here, this, this is, the, the threats are not new, the, the vulnerabilities, none of this is new, and if you have the security controls in place, that we've always had, or, you know, the last, like you said, half a decade, we've been recommending these things, and it's just making sure you have these controls in place.
19:10: So that when threats do arise, you're prepared, you're ready to address it, And you know, hopefully this whole experience of speeding things up, speeding up the, the pace of Devsecops, hopefully it's a, a lesson for all of us on best practices and what we need.
19:32: Well, it is also a lesson of, of, it's no longer sort of an aspirational thing to shift left, you know, especially as we move ourselves from a, Human driven SDLC and we're moving into an agen agent driven or agent controlled SDSC SDLC.
19:50: I think we're, we're gonna be in sort of in a hybrid situation for a long time to come.
19:55: you need to ask questions like, are we even grounding our agents right?
19:59: how would our agents know about, about, new security vulnerabilities or new supply chain attacks?
20:06: Here's a little rhetorical question for you, Andrew, not so rhetorical though, when is the, do you know when the knowledge cut-off date of chat TPT is?
20:16: The knowledge cut-off date of chat GPT, it's gotta be within the last 5 years, I would think.
20:24: It is, it is, it's in fact, you know, I believe the consumer versions, that most people tend to use, sit somewhere around, somewhere around late 2024.
20:35: So, you know, a little bit of a foundational architectural piece of piece of knowledge about just the LLMs that power these agents is that they naturally do not know anything past their knowledge cut-off date because they are trained up to data on that date.
20:53: So for example, if you ask Opus 4.7, which is the current, current bleeding edge model, of, of Anthropics loud as an example, that noise cut-off date sits somewhere around, early 2025.
21:08: So, and you can ask it, you can literally go and ask the model that question, the model will answer and say, here's my, here's the date cut-off that I have.
21:16: That represents a natural gap of, you know, both from a security perspective as well as a productivity perspective, because if you think about, think about, things like, hey, should I be using this version of open source, should I be, does this version of.
21:33: This MPN package, is it actually affected by a malware attack.
21:37: The foundation models right now have no way of knowing without grounding that information because they just have not been trained on that data at all.
21:45: So, one of the sort of key things to be thinking about is as you introduce these agents, it's not only just that you're gonna have to use them to produce, patches.
21:55: it's also that you're gonna have to think about how do we grant these models in the knowledge, knowledge of today, because the supply chain attacks really they are short-lived events.
22:05: They, they rely on the fact that you don't know about them yet.
22:08: They fly under the radar when they become known that that kills the taps.
22:11: So the sooner you can get the knowledge, and especially to your agents as they start, taking over some of these patching activities, the better it is because that means that you're able to execute things, at, at a faster clip.
22:24: That's probably one of the sort of biggest risks to these kind of programs right now is, is that gap is well-known, it's well understood, and there are no easy solutions except of course, what we can offer to organizations to help ground those agents.
22:38: Yeah, that's really interesting.
22:40: That's even more recent than I thought, you're saying, you know, end of 2024, beginning of 2025, that's the knowledge cut-off date for some of these models.
22:48: I, I was thinking it was even longer than that, I, you know, it, it feels like AI has been here for for longer than it has.
22:56: You, you, I was thinking, OK, maybe 2021, 2022, around there, but it really is so new, we're recording this in 2026.
23:06: And, yeah, the thought that it's really only been around for 12 years, that, that is very interesting, and, and, It, I, I like the point that you mentioned about this hybrid approach. 23:22: It's not fully agentic, we can't fully hand the reins off to AI at this point.
23:28: We need that human element that actually grounds the AI in the intelligence that we have as humans, because our knowledge cut-off date is much, much longer than 2024.
23:43: we, we go back quite a ways, so.
23:46: I, I, I like the, the hybrid approach, it's a good point that you make there.
23:50: Yeah, yeah, and I mean, it's, it's not only just about humans, but it's also about the fact that there's a lot of intelligence, specialist intelligence, takes a lot of time to gather the, the models just don't have that data by default.
24:01: so, and also that that landscape is evolving at all times.
24:06: The models also don't have the human tribal knowledge like human developers, we know where the CI system is, we know what best practices are, we know what not to do.
24:16: The models lack all of that organizational context, so to be able to also grant them on what not to do, is very important because that's how you avoid a lot of sort of reasoning loops, that's how you avoid a lot of long-term.
24:28: Patching issues, but also avoid the model falling foul of like a known ransomware attack package as an example.
24:35: Mhm, no, that's a great point.
24:37: well, Ilka, I know we're out of time here, but as we wrap up, I, I just wanted to ask you, do you have any final thoughts, any final recommendations for people around Mythos, or around AI in general?
24:52: Well, I think, I think, the mythos situation really it, you know, as we, as we discussed, is a case of more is more, you know, the speed of, all activities is going to speed up, your ability to produce code yourself is going to speed up, the ability for adversaries to be discovering.
25:08: New vulnerabilities and putting them into active exploitation is going to speed up, so sort of recognizing that that's where the puck is going, I think it's a really good and important exercise to be running now.
25:18: We've got about 30 to 90 days before the first vulnerability starts dropping.
25:24: not all of them are critical, mind you, not all the 3000 vulnerabilities they claim are gonna be at that severe critical range, so a lot of it is going to be, going to be less severe stuff, but asking those fundamental questions now of would we know where this would affect us, how would we deal with one a week, you know, one a day, even, you know, whatever that pace is going to be.
25:46: It, it affects the answers and it kind of forces you to forces you to think about some of the fundamentals.
25:51: Do we have the fundamentals in place?
25:52: Maybe we couldn't require our human developers to go through a centralist artifact repot to get, open source, but maybe we should do that for all of our agents if, if, if nothing else.
26:03: So asking those sort of basic questions now I think is, is very important and working with vendors like ourselves, who are specialists in the field will help, make that process go faster as well.
26:13: So don't, don't be a stranger.
26:15: I love it.
26:16: Well, thanks Silka again, thank you for your time.
26:19: thanks for sharing your thoughts on Mythos.
26:21: It's clear that you're an expert on the topic, and this is something that is very new to a lot of us.
26:29: So to have people like you who are very well versed on the subject, it's very helpful.
26:33: So, thank you for taking some time today and, to our viewers who are watching this episode, I wanna remind you to please subscribe if you're watching on YouTube.
26:44: or if you're listening on Spotify, please leave us a review, and we'll see you next time on another episode of Open Source Open Mic.
26:51: Thanks.
Related Resources
Mythos FAQs
What is Mythos?
Mythos is Anthropic's next-generation AI model designed to analyze source code and identify software vulnerabilities. According to reports discussed in this episode, Mythos has demonstrated the ability to discover previously unknown security flaws, including vulnerabilities that remained undetected for years.
Why is Mythos important to cybersecurity?
Mythos represents a major leap in AI-powered vulnerability discovery. It can help security teams identify and fix software flaws faster, but it also raises concerns that attackers could use similar AI capabilities to discover and exploit vulnerabilities at unprecedented speed.
How does Mythos find software vulnerabilities?
Will Mythos replace human security researchers?
No. While Mythos can automate vulnerability discovery and analysis, human expertise remains critical for prioritization, validation, remediation, and understanding business context. The future of cybersecurity will likely be a hybrid model combining AI capabilities with human oversight.
How will AI change software security?
AI is accelerating both software development and vulnerability discovery. Organizations can use AI to find and fix vulnerabilities faster, but they must also prepare for a significant increase in security disclosures, patching requirements, and software supply chain risks.
What is software supply chain security?
Software supply chain security focuses on protecting the open source components, third-party dependencies, and development tools used to build modern applications. Since most software relies heavily on open source, organizations need visibility into what components they use and whether those components contain vulnerabilities.
What are SBOMs and why do they matter?
A Software Bill of Materials (SBOM) is an inventory of all software components used in an application. SBOMs help organizations quickly identify whether they are affected when new vulnerabilities are discovered in open source dependencies.
How can organizations prepare for AI-driven vulnerability discovery?
Organizations should strengthen software supply chain security practices, maintain accurate inventories of dependencies, implement continuous monitoring, establish vulnerability remediation processes, and conduct regular incident response exercises.
Will AI increase the number of reported vulnerabilities?
Yes. Industry experts expect AI-powered tools like Mythos to dramatically increase the rate at which vulnerabilities are discovered. This means organizations will need better prioritization, automation, and remediation workflows to keep pace.
What are the key takeaways from this episode?
The rise of AI-powered security tools like Mythos is accelerating cybersecurity. Organizations should focus on software supply chain visibility, vulnerability management, SBOM adoption, AI-assisted remediation, and strong security fundamentals to prepare for the next generation of threats.