---
title: "AppSec Has Never Mattered More: The Time to Exploit Is Now Negative"
url: https://appsecpodcast.com/appsec-has-never-mattered-more-the-time-to-exploit-is-now-negative/
date: 2026-10-29
duration_seconds: 2888
season: 13
episode: 18
guests: ["Hannah Foxwell"]
topics: ["AI and LLM Security", "Vulnerabilities and Exploits"]
audio: https://www.buzzsprout.com/1730684/episodes/20031246-appsec-has-never-mattered-more-the-time-to-exploit-is-now-negative.mp3
video: https://www.youtube.com/watch?v=btK-TG3RhKs
transcript: true
---

# AppSec Has Never Mattered More: The Time to Exploit Is Now Negative

*October 29, 2026 · 48 min · Season 13, episode 18*

with [Hannah Foxwell](https://appsecpodcast.com/guests/hannah-foxwell/)

on [AI and LLM Security](https://appsecpodcast.com/topics/ai-security/), [Vulnerabilities and Exploits](https://appsecpodcast.com/topics/vulnerabilities/)

[Audio](https://www.buzzsprout.com/1730684/episodes/20031246-appsec-has-never-mattered-more-the-time-to-exploit-is-now-negative.mp3) · [Video](https://www.youtube.com/watch?v=btK-TG3RhKs)

## Show notes

When one AI model can find a vulnerability, another can write the exploit, and the gap between discovery and attack is collapsing. Hannah Foxwell, co-founder of BIMP and former product director for Snyk Container, joins Chris Romeo and Robert Hurlbut to argue that the time to exploit is now effectively negative, and that AppSec has never mattered more. Drawing on her path from release manager to DevOps, platform engineering, and container security, Hannah explains why security teams are facing an identity crisis, why a "dark software factory" with no human checks is a dangerous idea, and why so many AI security startups start with red teaming. The conversation also covers the industry's response to the flood of AI-discovered open source vulnerabilities, what leaders should measure beyond speed, and why empathy is still AppSec's most important tool.

This episode is sponsored by [Origin](https://appsecpodcast.com/origin). Let your agents work. Keep them in sight.

About Origin  
Origin is an endpoint AI observability platform that inventories, observes, and measures the usage of AI agents at scale. It gives organizations a record of what their agents do across endpoints, so security teams can monitor agent activity, investigate unexpected behavior, and enforce their policies.  
→ [Learn more about Origin](https://appsecpodcast.com/origin)

This episode is sponsored by [Corgea](https://appsecpodcast.com/corgea). Design it. Build it. Ship it. Corgea secures it.

About Corgea  
Corgea is an AI-native application security platform that secures software from design to production. It brings together security design reviews, AI SAST, dependency and IaC scanning, code quality checks, and autonomous pentesting—helping security and engineering teams find risk earlier, fix what matters, and ship securely.  
→ [Learn more about Corgea](https://appsecpodcast.com/corgea)

Connect with Hannah Foxwell:  
→ [Hannah Foxwell on LinkedIn](https://www.linkedin.com/in/hannah-foxwell/)  
→ [BIMP](https://bimp.ai/)

Mentioned in this episode:  
→ [IBM and Red Hat: Project Lightwell](https://newsroom.ibm.com/2026-05-28-ibm-and-red-hat-commit-5-billion-to-redefine-the-future-of-open-source-in-the-ai-era)  
→ [Chainguard's Athena coalition](https://www.chainguard.dev/athena)  
→ [Linux Foundation: Akrites](https://linuxfoundation.org/press/linux-foundation-and-industry-leaders-launch-akrites-to-defend-critical-open-source-software-against-ai-enabled-cyber-threats)  
→ [Shadow MCP Servers with Vandana Verma Sehgal](https://appsec.buzzsprout.com/1730684/episodes/19927672-shadow-mcp-servers-the-ai-security-risk-hiding-on-your-developers-laptops)

Chapters:  
00:00:00 - Cold open: security teams face an identity crisis  
00:00:25 - Meet Hannah Foxwell  
00:00:47 - Welcome and Hannah's origin story: from release manager to DevOps  
00:06:34 - AI as an architecture partner  
00:11:06 - From DevOps and platform engineering to AppSec  
00:13:43 - Have we reached AI transformation?  
00:15:21 - The dark software factory  
00:20:33 - AppSec's identity crisis and the vulnerability tsunami  
00:25:14 - Why AI security startups start with red teaming  
00:27:31 - Does exploitability still matter?  
00:29:07 - The time to exploit is now negative  
00:30:16 - AppSec's role when every developer codes with AI  
00:33:00 - Security training for agents and stale training data  
00:37:54 - Trust, MCP, and layered threats  
00:40:30 - Breaking the approval queue bottleneck  
00:43:16 - Measuring speed and safety  
00:45:41 - One behavior to change  
00:47:26 - Wrap up

## Transcript

*8,342 words · assemblyai*

**0:00 Hannah Foxwell:** It's a really, really tough time for security teams because there's a little bit of an identity crisis happening in the industry. Folks know that the job they do today won't exist for much longer because AI tools can do a lot of it for them, a lot of the research, just like discovery analysis, but they don't know what coming out the other side looks like.

**0:25 Chris Romeo:** Hannah Foxwell is the co-founder of BIMP, the Base Image Management Platform, and an engineering leader with more than a decade of experience. Her work brings together cybersecurity, platform engineering, and AI with a focus on the people behind the technology. She also created AI for the Rest of Us, a community helping people understand and confidently use AI. Hey folks, welcome to another episode of the Application Security Podcast. Chris Romeo here at the, I guess, Application Security Podcast control center. if we had a control center, but in the world of AI, do we even need one?

**1:01** That's the question.

**1:02 Chris Romeo:** Joined as always by my good friend, Robert Hurlbut, my compadre in application security, and we're ready to dig in again.

**1:11 Hannah Foxwell:** Yes.

**1:11 Chris Romeo:** And so, Hannah, we're going to jump right in here. And since your first— this is your first visit to the Application Security Podcast, we'd love to know your security origin story. I'm going to phrase it a little differently than I normally do, though. So, if your security career was a comic book, what would we find in episode 1? And would this be a hero origin story or a villain origin story?

**1:37 Hannah Foxwell:** Oh, I think, like, when you said you were gonna ask this question, I went immediately back to, actually, it's probably just like the origin story of how I came to be doing what I'm doing, because I started my career as a release manager. And this was waterfall software delivery, where the path to production was just fraught with unexpected challenges and problems. It's like every time you deployed what you thought was, okay, it's a release candidate build, it's finally ready, like 6 weeks late, whatever, like, let's try it, and it would just blow up in an environment. And it was because I was literally the human integration layer between the dev team, and the ops team, and the test team. Like, I came to really understand, like, the dynamics of, like, software development teams very, very early in my career. We're talking about the first couple of years of my career. And so, that's how I became interested in DevOps. And that's how I became interested in Agile and in continuous delivery, because I was like, how is this so hard? Like, how, why can't we take some software from this team and put it on a server over here and have it do what it's meant to do? And so, you kept digging and digging and digging. And one of, like, one of the many, many, like, challenges that I had as, as we kind of improved gradually our process around software delivery, was making sure that all of everyone could keep up. And so, while I was in this role, we went through the transition from waterfall development, like releases once every quarter, to more agile, like a potentially shippable build every sprint, every 2 weeks. And for us at the time, it sounds mad, doesn't it, saying it now? For us at the time, unthinkable. Unthinkable. It was like hard to imagine. And the security team hated me. They absolutely hate. So, when I say, is it hero or villain? Like, I was, I think I was doing good work, but to them, like, to the security team at the time, I was the villain. They were like, well, how on earth do you expect us to keep up if you want to deliver a new release every 2 weeks? And I'm like, doing my thing, like, well, let's figure it out together. Because like every release is much smaller than what it used to be. If we do a release every 2 weeks, it's nowhere near the magnitude of what our quarterly releases look like. So, what can we do in, like, this smaller and more continuous way? And so, yeah, I would say that I was, for sure, in their eyes, the villain. But I think overall in my career, I've tried to be the hero. I've tried to bring people with me instead of imposing change on them. And overall, hopefully, reducing the pain and risk that I used to experience, like, very, very vividly when I was that release manager trying to get the humans and the machines to do what we needed them to do for the business.

**4:26 Chris Romeo:** Yeah, my closest I think I ever came to being fired in the corporate world was a result of a tiff that I created with the release team. And so this was at a really big company. I've mentioned in the past, it doesn't matter now, but there were, it was a release, the release team wasn't 10 people, it was hundreds of people. And I was trying to get them to change something in the release process to better line up with what we needed from a security perspective. And I stepped on the wrong toes.

**5:02 Hannah Foxwell:** Oh my.

**5:03 Chris Romeo:** Luckily, luckily I had a, uh, the best manager I've ever worked for, a guy by the name of Tom Sweeney. And, uh, he, he had told me when I first started working for him, he said, I only have one, one kind of rule. And that is, if somebody's coming for you, you gotta tell me in advance. He said, I'll cover you. I'll, I'll, I'll do what I gotta do to protect you, but just don't let me get blindsided. And so that was the only time while I was working for him that I ever called him and I said, Tom, listen, they're coming for me. And he goes, I got you.

**5:37 Hannah Foxwell:** All of them. The enormous release team.

**5:39 Chris Romeo:** Yeah. See, he goes, I got you. Don't worry about it. And he got involved and we worked it out and we were all friends at the end. But, uh, they were literally calling for my head on a platter. at one point during this negotiation. So, that was my release experience.

**5:54 Hannah Foxwell:** Well, and I mean, there was so much, like going through these transitions, through like agile transformation, through DevOps transformation, trying to do things differently. You know, it can be quite threatening to folks, like when what they know is their job, the thing that they are proud of, the thing that they're good at, that they built a career around, starts to feel under threat by this newness. And I think there's probably a lot of people who feel that way about AI today, but there's only one thing that you can really be sure of is if you want a career in tech is like change is constant and you can't really, like none of us can get really precious about like, oh, I just want to do this one thing forever and ever.

**6:32 Chris Romeo:** And I mean, we are in an era of change that I've never seen in my 30-year career. things are literally day-to-day right now. Like, you can look on X and go, oh, that's kind of cool. I mean, I did, I did it this morning. I found a skill that generates an architecture diagram from your entire repo.

**6:54 Hannah Foxwell:** Yeah.

**6:55 Chris Romeo:** And it just blew my mind. Like, as somebody who likes to architect things from a security and privacy perspective, all of a sudden, I've got this diagram that was generated for me from the repo. You click on one of the boxes, and it literally lights up with, with motion to show you, like, where the is moving. It was like my first ever interactive data flow diagram. I'm like, why didn't I have this before? Like, I need little motions to, you know, entertain me while I'm doing threat modeling, for example. All right, Robert.

**7:27 Hannah Foxwell:** Yeah, no, absolutely. Like visualization and like making the human interaction with our code and with our technology, it's an area that I'm very interested in because interacting just through a chat window I don't think it helps us go as far as we could if we had ways of reasoning about it in a more visual way. So, I love that. You'll have to put the link to that skill.

**7:50 Chris Romeo:** I'll have to put it in the show notes here. I just wrote about the dangers of skills on LinkedIn. So, that's a whole other thing. But, and then I violated my own guidance not 12 hours later. I'm like, wow, that's cool. And I didn't even do a code review of it or anything like I should have. But yeah, I mean, I'm finding, Hannah, as well, like to your point, The interaction with AI, if you keep it just to the conversation, is not as valuable as what I'm finding using the AI to interact with the architecture and the design early and guide it via architecture and design versus trying to guide it command by command through a chat window or something. So kind of forcing it, it and myself to pull back, look at the 20,000-foot 10,000-meter, I'm guessing, view, you know, and then, and then go back down and, and dive at various levels of abstraction like specs and things before letting it code.

**8:50 Hannah Foxwell:** Yeah.

**8:51 Chris Romeo:** And I think the danger is, is people are still just letting it rip right now. And I found the first thing I built where I let it rip, I literally deleted the whole thing. I said, I got to start over because this is, this is impossible to follow. And then I started kind of walking through the various layers of architecture structure. And then I started to get the results I was looking for. Like things were actually working according to what I was looking for.

**9:17 Hannah Foxwell:** Yeah. That kind of like one-shot mega prompt kind of like building is okay if you just want to test like an idea and like, and you will, you know, you're gonna throw it out and you're gonna like have to chop it up into smaller things that you could be much more directive over. And it's like, well, let's make sure this slice is right before we move on and build on top of it. So, yes, it can be useful for helping you make informed decisions. It's not all waste. I mean, again, it's another completely different dynamic in the world of software development, where every single minute of engineers' time was just fought over. Like, what's the highest priority? Doing the backlog shuffle. And now, we're much, much more able to just test an idea in actual software. not in like a design, like, um, a wireframe or something like that. We can actually just test an idea, see if it's worthy of, of investing time in, um, which I think is fabulous. It just helps us make better decisions.

**10:16 Chris Romeo:** This episode is brought to you by Origin. Quick question, would you hire 1,000 people without knowing what they can access or what they're doing all day? Of course not. But that's how most companies roll out AI agents today. You get the prompt and the final answer, Everything in between is gone when the terminal closes. Origin finds the agents on your endpoints and rebuilds their work. Every command, every file, every handoff between agents on one timeline. Origin endpoint AI observability. See what they found in their own traces at originhq.com/appsec. That's originhq.com/appsec.

**11:08** Definitely. So, Hannah, given your background in DevOps and platform engineering, how would you think that is an advantage perhaps over somebody who came up in cyber?

**11:16 Hannah Foxwell:** Well, I think it's a little bit of both. And so, going through the sort of DevOps transformations and adopting continuous delivery, and then I was platform engineering director at VMware. And so, we went and we helped our customers adopt these practices for the first time, you know, really, really coaching them through some of that quite, like, quite challenging change facilitated by technology, but in a large part about people and process and culture and trust as well. And I think, you know, those things, those things helped me step sideways quite seamlessly into cybersecurity when I went and joined Snyk. So, I was product director for Snyk Container, and it was all of the same things, really. It was all of the, it was all like, well, how do we know that we can trust these container images? How are we going to reduce the cycle time? How are we gonna build those feedback loops in so that, you know, vulnerabilities can be remediated with lower friction and also shorter cycle times? And so, I think, you know, a lot of the practices that I'd been preaching for like over a decade in my career, they just transferred really seamlessly over. I think one of the, one of the things that, you know, I felt very, very strongly when I, when I joined Snyk as a, as as a product person, was this like new empathy for security people. And like understanding that like the information overload that folks receive, like the number of bad news dashboards that don't always agree with each other about what flavor of bad news you've gotta look at first. And it's, I was just like, well, how on earth can we cut through all of that noise and build something that's much more like user-centric? Like, what does a user need? What does the security person need at this moment in time? Or what does the developer need at this moment in time? And trying to bring those two together. And so, I think, you know, in platform engineering, we would always beat the drum of like building your platform like it's an internal product. And then I was given this amazing opportunity of building a product that was designed for developers and security people. But that, it for sure opened my eyes about how challenging the role of application security teams is. And yeah, like, it's still a challenge today. In fact, it's a greater challenge today because there's a hell of a lot going on with new challenges that AI is bringing with it, especially this year.

**13:42 Chris Romeo:** You mentioned DevOps transformation as a concept. Have we reached AI transformation yet, or are we still in our— is it, are we still in too much of an infancy stage to even call it AI transformation?

**13:58 Hannah Foxwell:** Oh, you know, like one of my friends runs a consultancy around this. So, I'm gonna borrow this language from them. And he talks about the ladder in AI transformation. And I think most people are quite comfortable going from, I type code by hand, to, I prompt an agent in a small way, like every task I partner with an agent on. And then, I think that there's this like, growing sort of layering of capabilities. It's like, well, how agentic is your software development? Are you doing spec-driven development? Are you using loops to make sure that the agents are checking their own homework? Do you use like an adversarial review with another agent and another model checking that? And I think there's this layering of capabilities and teams are operating at different levels. You know, people are talking about the dark factory now, aren't they? No one ever looks at the code. Like, that's apparently the goal. We're never gonna look at code ever again. And some teams, like, if you believe the marketing, are genuinely operating in that sort of dark factory kind of mode now with their software development, but that's not everyone. So, I think when you talk about transformation, it's not really done. I think we are all at sort of different phases of adopting AI in software development. And I think the destination may not look the same for everyone. Because it doesn't have to.

**15:20 Chris Romeo:** I think the idea of the dark software factory is a very dangerous concept. And I think we're, as an industry, we're becoming victims of the bleeding edge. And what I mean by that is there's like 5 or 6 companies that everybody wants to emulate today. It's Cursor and Socket and Anthropic and OpenAI. And there's this kind of grouping of companies that everybody says, well, that's how we should be doing it. That's how we should strive for. But like, you think about a bank, a bank is not going to an agent-full, human-less software factory probably ever, just because it's a risk management nightmare. And maybe 25, 30 years in the future where you have all the ability to have all the evidence and all the right checkers, But they're still gonna have a human in the loop in like a large financial organization. Yeah.

**16:22 Hannah Foxwell:** Well, it's the 4 eyes as well. It's not just one human, it's like, it's 2 and probably a swarm of agents. But yeah, I agree with you there. I agree with you there, 'cause I will admit that, you know, I was of the opinion for a while that, you know, everybody should be doing continuous delivery, and like every build should like just automatically go through automated tests, propagate its way to production. And I think actually, it's all about how much risk tolerance your business has. It's all about the business at the end of the day, how much risk tolerance your business has, what the consequences of a failure are, like how expensive is it? And, you know, it's like what capability you have in your team, because actually to build a system like that, that can, like, you know, whether it be agents or in like the old-fashioned way with automated tests and things like that, to build a system and a platform that's capable of that is an investment. Not everyone has that much money to invest in that particular problem space. You might wanna invest your money in solving problems for your end users and your customers. Like, you may not even be selling software. You might just be using software to enable your business. So, yeah, I do think that there is, There's room for one version of right. And I've come to learn that over, over many, many years of like kind of beating the drum about one way and then realizing that, you know, that was, that was not, not necessarily the right thing to do.

**17:48 Chris Romeo:** I guess a good way to think about it, there's more than one definition of right. I think that's a good takeaway that we just have to remind ourselves often because there are going to be some bleeding edge AI companies who are able to truly go to a full-on software factory with no human intervention or limited human intervention? The security person in me is building a stain.

**18:12 Hannah Foxwell:** Yeah, I mean, the interesting thing about those places is they're spending a huge amount on software development. They're not doing this to save money, for sure. They're hiring like the best agentic engineers who are building custom harnesses. They're taking more of a platform engineering approach to software development, but They're also burning through so many tokens as well, like, to build these things. And so, I think it would be foolish to look at these organizations and go like, well, they must be, like, because developers aren't typing code or even looking at code, that they must be more efficient and spending less. And I don't think that's necessarily true.

**18:54 Chris Romeo:** I can't wait for the metrics studies to actually come out about this.

**18:57 Hannah Foxwell:** Oh yeah. Will people be honest?

**18:59 Chris Romeo:** Like, I, you know, You should be able to look at the data and extract it though, whether the people are honest or not about impact and of true efficiency based on defects generated and things like that. Are we better or are we worse? That's the study that I want to see and I haven't seen it put down yet.

**19:20 Hannah Foxwell:** No, no. And I do, you see it in like the salary bands and things like that at the moment though, like the people who are building the thing that builds the thing are commanding quite enormous salaries at the moment. And the folks who are just, just, just, I don't want to use that word, the people who are building the thing, like traditional product engineering, are maybe losing out in the job market compared to these enormous, enormous packages that are going from places like you say, Anthropic and Kirsten and folks like that.

**19:56 Chris Romeo:** Who thought developer would be a 7-figure opportunity, or senior developer?

**20:02 Hannah Foxwell:** I know. I know. I wish someone, I could go back in time and someone put a little word in my ear, because when I was first like a trainee, I was taught COBOL. And I was like, yeah, not sure it's for me. I'm going to go this other route, which is how I ended up more in management and delivery and product. But, uh, perhaps I should have stuck with the programming a little longer.

**20:27 Chris Romeo:** Perhaps we all should have. This security thing, maybe we, maybe it was a miss. So, Hannah, I know you've been doing some user research lately and you've been hearing from AppSec teams. Like, what are you, what are you hearing from folks in the field? And really, what's, what surprises you based on that analysis?

**20:46 Hannah Foxwell:** Yeah. So, um, at the moment I'm working on a new product. Um, I spent, I spent last year advising a ton of startups and helping them understand like a lot about AI, but also platform engineering and DevOps and product leadership. And this year, I decided to build my own thing. I was like, well, why am I helping other people? I should just help myself. So, it's called BIMP, the product. It's still in closed beta, so you can't go buy it yet, but please follow me and I'll let you know when you can. Basically, it's a supply chain security product because I became fascinated with this, this challenge of like, actually, how do we know where our software comes from and how do we know whether it's good and how do we know where it ends up? And it's a vast, vast problem space. And I went, I'm doing tons of user research with application security teams at the moment. And like, it's a really, really tough time for security teams because There's a little bit of an identity crisis happening in the industry. Folks know that the job they do today won't exist for much longer because AI tools can do a lot of it for them, a lot of the research, discovery, analysis, but they don't know what coming out the other side looks like. So, there's a little bit of an identity crisis, but at the same time, there's an actual security crisis happening whereby we have— Like these incredibly powerful new models that are discovering vulnerabilities at such a, like such an enormously accelerated rate, like latent vulnerabilities that have existed in the open source software that we all run as part of the infrastructure stacks, like everywhere. And it's had a humongous reaction in the industry. You know, you've got IBM and Red Hat throwing $5 billion helping to solve the problem. And ChainGuard with the sort of Athena coalition going out there and being like, right, we'll be a clearinghouse. And the Linux Foundation standing up an emergency project called, like, another clearinghouse called Acritys to help coordinate, like, a more sustainable, like, response to all of this. And it's like, we've got this tsunami of vulnerabilities coming at us. And yet the thing that I hear from the application security teams is, am I relevant? And it's like, and you've never been more relevant, my friends. Like, we've never needed you more. And so, And so, yeah, like I think again, it's like, it's something that I care about is like helping these folks cut through the noise. Like that's what, that's what I'm all about at the moment because the noise is only getting louder and louder. And there's new threats posed by AI agents. The risk tolerance level, I believe, is also shifting where I used to talk to folks and they would only really concern themselves with like critical vulnerabilities. Now that you have agents able to chain together lower-level vulnerabilities and create an exploit that's far more powerful, I think the threshold by which, you know, the risk tolerance level has shifted as well. And all of this just creates volume, volume for the same few people in every organization that have to manage this risk. And so, yeah. Like I would say, like every team I talk to, I say like, one, you're more relevant than ever. We need you more than ever. And like, two, how can I help you cut through all of this so that you can actually do good work. Because again, like, we are all figuring this out. What you did yesterday isn't going to serve you well tomorrow, but we don't know what tomorrow looks like necessarily. So, yeah, that's like—

**24:24 Chris Romeo:** That's a great word for our application security folks who are listening to this that are maybe struggling with that right now, is you are more relevant now than you've ever been. This episode's sponsored by Corgia. Design it, build it, ship it. Corgia secures it. Corgia is an AI-native app security platform covering your whole software lifecycle, from the first architecture diagram to code in production. It brings design reviews, AI-powered code scanning, and autonomous pen testing into one platform, so your team spends less time chasing noise and more time fixing security issues. what matters. Visit corgea.com, C-O-R-G-E-A.com. Corgia, one security platform for the entire SDLC. I'm curious, Hannah, based on your kind of startup advisories over the last year, you know, before you got into building your own startup, when I look across the AI security startup landscape right now, Seems like there's a bulk of people focused on the red teaming side. And I think that's a common theme as tech, as transformations happen across industries and security is somehow involved. People always lean towards the red team side first. But I'm just curious to get your take on that. Like, are you seeing that across the startup scene right now?

**25:55 Hannah Foxwell:** And—

**25:55 Chris Romeo:** Why do you think folks start with the red team side?

**25:58 Hannah Foxwell:** I mean, I think the red teaming is a really, really helpful tool because basically these are the— this is the toolkit that's now like available to the hackers, your adversaries, the people who are like— it's like the baddies and the goodies, isn't it? So, you have to really understand the capabilities there to be able to defend well against them. So, I think in many ways, the red teaming is a lot of folks going like, I need to understand this better. And, oh, I've learned some stuff, and, oh, wouldn't it be great to like build a business around all of this knowledge that I've acquired? And so, I think there's a little bit of that. I think also, like, teams would be foolish not to turn these tools over and like test their own systems themselves and see what like perhaps a novice could achieve with a powerful LLM at their fingertips, because it's a lot, actually. It's a lot. And so, yeah, I think that's, I think, I'm seeing that as well. I'm definitely seeing that as well. But I think, like I said, I think the risk tolerance level of a lot of organizations needs to shift. And I think that's generally the outcome that we're going to see people sort of navigate their way towards. It's like, oh, well, we need to, this acceptable risk is no longer acceptable. And from there, there will be re-engineering and re-architecture and investments. But I think a lot of folks haven't got to that point where they've even discovered like how, where the gaps are and where they need to change and invest.

**27:30 Chris Romeo:** And one of the other things that I'm finding as I look at some of these more red team-focused AI approaches, they seem to be funneling in on this idea that the exploit is a really important piece of this. And when I think about this as somebody who's been building things for a long time and part of security organizations at various companies. I just don't know that I care that much if there's an exploit possible, because when we get into that state, okay, I don't want to add assurance based on the fact that there's no exploit available. Because when you think about the difference between, I haven't used Mythos, but what I've heard Mythos can do versus Fable or Opus or— Astra, it doesn't matter. I don't, I don't care which model provider we're talking about, but we're talking about like capabilities of model. So just because an exploit couldn't be generated by one doesn't mean the other one couldn't generate it if it got loose.

**28:29 Hannah Foxwell:** Yeah.

**28:29 Chris Romeo:** It also doesn't mean like in the old days, the fact that an exploit wasn't available today wasn't a reason not to fix something because a new attacker could come up with a new class of attacks and dream up some new scheme that nobody had ever thought of 6 months in the future. And all of a sudden we're not, we didn't fix it because we didn't think there was an exploit available. Like, I think there's some real tension there, and I don't know that I care if there's an exploit available. If there's an identified flaw in how the software has been designed or built, I say we just fix it and not worry about the exploit. But I mean, what are you thinking about this?

**29:06 Hannah Foxwell:** Yeah, and you know, they're like, um, like, it's— some of it is a bit doomy and hypey, but I do believe that Like the process of responsible disclosure around CVEs, it's a very human-speed process. It's a coordination process. But if one LLM can discover the vulnerability, then like you said, the other one can discover the exploit. And so, the time to exploit is outpacing, it's almost negative. Like before the CVE is even like out there declared with a number, with a remediation available, with a patch available, The exploit has been available for a while, probably, like if you went looking for it. And so, I think then we have to look at defense in depth. We have to look at the ways we can make sure that those failures don't matter or that we contain the blast radius of, like, what is the worst-case scenario here and how can we build multiple layers of defense around it? And that's really, like I said, this is where people are, like, folks are going to do the threat modeling again with a much, much lower risk tolerance. That's what I truly believe.

**30:15** Let's revisit a moment about developers using AI coding tools typically every day. How does that responsibility for secure code changes? Where does that lie? Where does it land? And has anyone actually agreed on that?

**30:29 Hannah Foxwell:** I mean, before AI, had we even agreed on it? Like, let's go back to basics. And I do think that building those feedback loops and safety nets into your software development lifecycle, that hasn't changed. That hasn't changed. We still need to be scanning. We still need to be finding these common sort of patterns of failure. The one thing that I do think that I would recommend for teams, and again, it's AI agnostic, is think about the role that security plays in your organization. Is like, I know it manifests itself as the sort of a source of no and a source of friction. Like, And folks in security get known as like the police, and it's like, okay, well, I've checked, and actually, no, you're not good enough yet, and like, we need to fix this now. And it can create like this conflict, this internal conflict by teams who want to go fast, and they now want to go faster than ever with AI, and teams that are just like, well, actually, like, we do have to go fast, but safe, like, we have to go safely. And so, 2 patterns that I've seen work in the real world are security teams operating in more of a DevSecOps, or security engineering, or platform engineering mindset. So, building the solutions that automate part of the process, getting the humans off that, whether that be like automatic remediations through some of the tools you bring in, or whether that be just building security into the platform layer so that vulnerabilities matter less. Teams who take that sort of engineering-focused approach, building a service that developers can consume from you, can scale their impact much further than a team that are trying to police what other people are building. The other thing that I would always advocate for is education and enablement and security being a place where people can come for advice without judgment, and also outreach, just providing that education during onboarding or periodically anyway. And so, I think those are 2 patterns that I've seen become very powerful, and they help to break down that source of friction that we so often see between developers and security. It's like the speed versus safety debate. It's like, no, no, it's both. It's both. We can have nice things, but we just need to think about how we deliver them in a different way.

**32:59 Chris Romeo:** I was having a conversation with somebody yesterday, And we were joking about, do we need to do security training for agents? Like, should that be, should that be a product offering? And like, it's not, it's, it's not that you need the agent to, you want the agent to absorb the content that we would normally try to impart to developers about security, but it seems like it should be a lot easier to do because agents are built to absorb large amounts of information quickly. And then need, they need our help to decipher how to store them in the right place so they become part of a pattern of behavior.

**33:39 Hannah Foxwell:** Yeah.

**33:40 Chris Romeo:** But do we need security training for agents? I don't know. Maybe that's a, that's a new idea.

**33:44 Hannah Foxwell:** No, it's just, it's just, it's another standard that you, like, if you don't provide that context, then the agent just works off their internalized weights and memory, like it goes off the training data and they could never know what your organization wants or needs. And so, I think it's not a bad idea. I know some teams who are doing the security review on the sort of pull request level, like the security review has gone through, like, well, let's look at the nature of this diff, this pull request. Like, is it high impact from a security point of view? If so, let's go and analyze it further. Let's see if we can find issues with it. And so, There are ways of integrating like some of your organizational standards into the context for agents in a way that doesn't add too much more friction. But yes, like you gotta be careful. Like you don't want to bloat every single agent interaction with every single thing they might need to know. I think it's much more powerful to build in feedback loops that automate some of the safety nets.

**34:52 Chris Romeo:** Yeah, nobody's got the budget to include all the security guidance we want wired into every transaction.

**34:58 Hannah Foxwell:** No, you can't.

**34:59 Chris Romeo:** You can't. One of the things that you just mentioned was about kind of model training, and I wanted to share something that I've noticed over the last week or two that I didn't notice when using a more general-purpose platform. So I've switched from using macOS to Linux for my development environment. And of course with Linux, it's so everything's command line driven. So you can have an agent basically do all the work that I used to do as a sysadmin and spend 2 days doing in 5 minutes. But I've noticed a pattern that I never saw before with macOS in that the training data is a lot more restricted to a time in the past.

**35:45 Hannah Foxwell:** Oh, interesting.

**35:46 Chris Romeo:** So, It often struggles with like configuration of something. And I have to go back and say, go search the internet and figure out how this really works. And then it comes back and says, oh, you're right. It wasn't in my training data that this is how this configuration pane works now. But it was, it, it hit me a lot harder in that environment than I'd ever seen with macOS. 'Cause I can't say that I'd ever hit something with macOS that, that the current model didn't have some context on.

**36:14 Hannah Foxwell:** Yeah.

**36:15 Chris Romeo:** And some ability to give me a pointer or some advice for something. But when I, with Linux and things that are a lot more obscure, and it just caused me to realize maybe I put too much trust in the current training data, current context in the model. And there's a lesson in there from a security perspective as well. Like, don't get too comfortable with these things as they appear to be doing the things that you used to do as well or better than you did. Because the connections may not be behind the scenes like they look on the front end.

**36:48 Hannah Foxwell:** Yeah. And you have to think about the reality behind the scenes, even at the frontier labs who have so, so much money to invest in the training data and creating that for the new models, is everyone in product development's prioritizing all the time. It's like, do we do this or do we do this? Is it more important here or there? How many people are on Mac? How many people are on Linux? And so, yeah, if you, and I mean, it could end up with a place of a real kind of standardization of sort of our infrastructure stacks and our application stacks, because if the LLMs are really good at building certain things and really poor at building others, then those others— are inevitably gonna be harder for organizations to adopt. So, there's real leverage there in like getting, you know, getting your data into a place where it can be easily consumed by an LLM, because otherwise your users are gonna be held back by all of those like little paper cuts and friction points that you just described, Chris.

**37:53 Chris Romeo:** And then there's the whole trust issue of if you're getting that information somewhere, how do I know the LLM's getting a piece of information that is trusted from the manufacturer and not an attacker's diversion.

**38:06 Hannah Foxwell:** Yes.

**38:07 Chris Romeo:** That's what I find with AI right now. There are so many, every time I talk about something, a new threat pops into my head of like, it's, I haven't reached any, anything where I'm like, this is solid. I can't, you know, the security person in me can't find anything to pick apart. Like everything's got some number of layered threats. that are just part of how it's been built. You know, we just did an interview, a previous episode with Vandana from OWASP talking about MCP. And it's like MCP was the same thing. It's like every time you turn a corner, it's a great, it's a great protocol, does a lot of great work to make things possible. But there's, did it, did it happen so fast?

**38:49 Hannah Foxwell:** Yeah.

**38:49 Chris Romeo:** That it didn't get what it needed. So.

**38:51 Hannah Foxwell:** Well, there's, there's so, there's so little control that you have As a user as well. If you give your agent access to the internet, which it needs for a lot of things, like, you know, just to go look at docs for stuff and things, if you give it access to the internet and it ends up at the wrong URL and that URL has been hijacked, then you can very quickly end up with prompt injection and context poisoning very, very rapidly. And you've got no idea that that happened because it may not have been visible to you at all. And so, and so, yeah, I don't know the answer to that, but I'm the same. Some, like, sometimes I think about this for a bit too long and I have a little, I Crisis, like shut my laptop and put on my hiking boots and go and stare at, like, stare at something natural for a while. Like, oh God, technology's too hard.

**39:33 Chris Romeo:** That is a good idea.

**39:35 Hannah Foxwell:** Yeah.

**39:36 Chris Romeo:** I can say I've been in the same spot where I'm like, I just need to be done for tonight because I've reached a point where I don't know if this is even possible or if we should go forward in this way because of some philosophical issues. So we're all philosophers now. I'm going to change my title to AI philosopher.

**39:53 Hannah Foxwell:** Yeah, I'm going to tell you how bad I've got. So I spend all day thinking about AI, talking about AI. I had a dream last night that I was a robot and I was like, wait, I'm not human after all. I'm some kind of sentient algorithm.

**40:13 Chris Romeo:** That's what Elon claims.

**40:18 Hannah Foxwell:** Exactly.

**40:18 Chris Romeo:** Yeah.

**40:18 Hannah Foxwell:** Anyway, that's where my mind went last night.

**40:21 Chris Romeo:** Terrifying. All right. Well, how about the approval queue? We've— in the world of AppSec. Yes, we're back to the real world with like, real people. Yeah, let's get back to— I would love to, I'd love to spend an hour or two talking about philosophy, but we got practitioners on this, on this record, or they want to listen to this recording. So, AppSec approval queue, what do we do with this in the AI world? Like, how do we not make it a stop sign, a limiting factor for how we can be effective with agentic AI?

**40:52 Hannah Foxwell:** Yeah. Everywhere you look, there's sort of new bottlenecks or shifting bottlenecks arising in the software development lifecycle. I do think I'm gonna be one of those people to say that AI can help to a certain degree, cut through the noise and make sure that the people, the humans who are managing that work, see the most important things more quickly and with like helpful information about how to, like what the next steps might be. I know that there's an AI company over here in the UK called Synthesia that do voice AI, and they've done some great writing about where they've deployed agentic AI into their AppSec team. to make them more efficient, to basically give them some time back so that they can focus on the more important things. And I think that's a model that we might see more people follow. I also mentioned, you know, having a sort of adversarial review by a separate agent that is put in the security reviewer kind of persona on your pull requests before a human ever has to look at them. So, there's a feedback loop there, potentially even an agent-to-agent feedback loop. So, the, you know, the pull request with the security flaws never gets even looked at by a human. It goes straight back to the agent that wrote it. So, yeah, there's things that we can do there. I don't think removing the checks and balances is the right thing. It's certainly not the right thing for certain organizations, like if you're regulated, like we talked about banks and finance earlier. The one thing that I have seen from a few different angles is folks trying to remove duplication. On their software development lifecycle as well, because if you keep all of your human review feedback loops and you add a load of AI reviews, you end up doing a lot of the same thing multiple times. And so, again, it's about your risk appetite. It's about proving these systems before you YOLO a new release into production for your customers to feel the impact of. I think it's about maybe continuously chipping away at those bottlenecks in your process. and seeing how you can really optimize around them without that, well, proving that you haven't compromised on the quality. That's what I would say. Like, take a data-driven approach, prove that you haven't compromised on the quality, and then try the next thing, and then try the next thing.

**43:15** You know, most of what leaders are measuring today captures speed, but what should they measure to know whether AI is helping developers build more securely? And would they act on it if the number looked bad? What are your thoughts?

**43:28 Chris Romeo:** Not a dashboard. Please, not a developer dashboard, which results in the bottom 5%.

**43:37 Hannah Foxwell:** Yeah. So, I read an awesome report by the Cloud Security Alliance the other day about some of the measures for success around vulnerability operations, like Volnops, like the vulnerability lifecycle. And one of them was cycle time, like from disclosure through to remediation. But another one was about like system saturation, like if you've got more like signals and risk entering the system than that you've got popping out the other side, then that backlog starts to grow. So, managing the capacity, like the saturation of the system, were— I hadn't seen it framed in that way before. And when I saw it, I was like, yes, like this, this is a wise metric because like actually your cycle time might be looking fine, but your system overall might not be. Or you might be processing only the top tier of things and you might be doing it quickly, but everything else is just lagging behind. So, yes, I think, again, like I always say, like speed actually requires safety. Like, you can't go fast unless you can go safely. And so, it's about building that speed, but in the right way, because otherwise, ultimately, you're gonna end up impacting your end users and your customers, whether that be through security issues or reliability issues. And those things are company killers. And we've seen it, we've seen so many high-profile companies have very high-profile outages because the volume of change moving through their systems and into production, it makes it very, very difficult to triage and fix, like, one of those, like, one in a million defects that's, like, 100 levels down. So, yeah, I think I would say that, like, looking at cycle time, like, things like MTTR, but also saturation, saturation of your system. Have we moved too much change through this system at the same time? and finding out where the balance is between speed and safety, not just measuring the speed side of things.

**45:40 Chris Romeo:** So, just to, as we start to wrap up here, Hannah, what's one behavior that you would want to see developers and AppSec people adopt as AI becomes the norm for how we build software?

**45:56 Hannah Foxwell:** I mean, I'm gonna cheat and give you 2 answers. I wanna cheat and give you 2 answers because I think that The thing that I think will set folks up for success as an individual navigating this change is curiosity. Actually just being very open-minded, being a bit of a sponge and figuring out new experiments to try and being a little bit bold about what you try because there's nothing to really stop you. the learning opportunity right now is enormous. So, I would say curiosity, but I also think everything that I've done in my career, the things that go well are the things where you bring empathy for the other people involved. And so, developers are having a bit of an identity crisis and their world has changed. On the AppSec side, it's exactly the same. And folks are wondering what the future looks like and having a bit of empathy and going back to basics and being like, well, we need to work together to solve a problem. Let's work together to solve the problem in a way that works for both of us. And it's been true with every technology transformation I've ever been through. It's like the folks that navigated it elegantly, shall we say, are the folks who did so with empathy and without judgment and without forcing folks to do things that they weren't comfortable with. So, I think having empathy in both directions at the moment is a very powerful tool.

**47:25 Chris Romeo:** Well, Hannah, thank you so much for joining us to be a part of the show.

**47:29 Hannah Foxwell:** Thank you for having me.

**47:32 Chris Romeo:** Yeah, I mean, you're— this has been a really fun exercise from philosophy to practical. We've crossed that entire journey from both of those concepts and gone back and forth a number of times. So it's a lot of fun. So good luck to you with BIMp, and we look forward to having you back again for a future conversation to just keep diving into how this AI thing's changing the world.

**47:55 Hannah Foxwell:** Yeah, absolutely. Relentlessly.

**47:58 Chris Romeo:** Thanks for listening to the Application Security Podcast. If you enjoyed this episode, subscribe, leave us a review, and share it with a friend or colleague who would enjoy the conversation. We'll see you next time.

---

Source: https://appsecpodcast.com/appsec-has-never-mattered-more-the-time-to-exploit-is-now-negative/