Jason Haddix welcome to the Noise 2 Signal Podcast. It's such an honor to have you on.
Thank you so much for having me. Appreciate it.
Jason, earlier this year you wrote an epic article, DEF CON is dead, security is cooked, and you had a question mark at the end of it. So I want to start the interview there. Was that sarcastic or did you mean it?
No, no, I think I think I was very frustrated by the experience of the chatter around the conference this year and how everybody was saying that narrative, but it was definitely a sarcastic title to me. I think that we're in a great position as as security people and things are changing obviously very rapidly and we we have to change with the times. We've had times where we've had to change before. and so that's the kind of thing I explored in the article was you know, like how we adapt, right? But
You know, we've been through this age of security's dead, pen testing's dead, you appsec is dead, cloud is dead. We've been through it eight billion times. And I don't think that AI is is much different.
I remember I remember when Metasport came out, the commentary was pen testers are dead. Then when the cloud infrastructure was picking up, the infrastructure jobs are dead. And when Claude Claude security came out, cybersecurity is dead. Like you know, sh you know, CrowdStrike is dead, Palo Aldo the there these guys have no moat.
Yeah. Yeah.
Mm.
they should just shut shut the companies down, give back the money to the shareholders. all kinds of random commentary was coming out in those days. Thankfully that commentary has now changed. and those those companies have gone to higher heights since then. And this is this is a pattern that I've seen over and over again. When web apps came out, web security was it fall well, fair web security followed.
Yeah.
Yeah. Yeah.
When cloud came out, cloud security followed. When mobile came out, mobile security followed. I feel like with AI, AI security is gonna follow up. So this is this is a pattern that keeps repeating repeating itself. what what do you think?
Yeah, I mean, I think that not only does AI represent a new attack surface, which creates a whole new security category for us, right? We're already seeing that today, right? Agents are very capable in doing offensive security and defensive security tasks, but they're also not super controllable at the moment. And then everybody who's baking in AI into their applications as features is having trouble with prompt injection and agent autonomy and cross-tenant data access and
We still haven't figured these things out. Now there these roots actually lie in some of these eras that you just talked about, right? you know, sandbox isolation, you know, monitor, you know, logging and monitoring, you know, detection engineering, you know, all this stuff is things that, you know, privilege escalation, you know, on you know, Linux distros, all this stuff has been around for a long time. It's just getting wrapped again in a new need, which is great.
I I do think that, you know, one of the arguments against that, you know, like security is cooked in general, is that AI replaces not just one thing that we've had, and and it doesn't add one thing. It adds a whole it impacts a whole domain of of skills, right? And I I do think that, you know, AI does have a larger sized impact. But I I gotta tell you, at least from my experience and my experience with the bug bounty community, which I have been
very involved with for many years. I'm seeing the best hackers only get more efficient. I'm seeing the best consultancies only release better products faster that are really moving the needle, ourselves included. and I am seeing no lack of news articles about hacks happening, you know, even at some of the model companies, right? Who should have this all figured out. If the if security's figured out, they should have it all figured out, right?
They probably have the worst security g given the example we have seen with Hugging Face and all those things. There's one follow up I want to make and the thing that you didn't mention is you know, obviously all these improvements are happening. but the thing that you didn't mention is this cost of AI is also control uncontrollable. Once you set these agents loose, they keep thinking, they keep thinking. You have to bring you have to bring them back into you have to ground them pretty quickly.
Yeah. Yeah.
yeah. Yeah.
Yeah.
Yeah, yeah.
And this is where the experience comes in. If you know what you're doing, if you know how to extract, if you know what what good looks like and what you can extract out of these systems, you can limit the use of tokens and still get and still get incredible value. The point, the point I want to make, and you mentioned that AI is having a broader impact, maybe a broader domain impact. And there is this false narrative that AI is end-to-end.
AI is end-to-end, you let these things you let these things do, and they can automate jobs, they can automate products. None of these products are no longer required. And I feel I I agree with the there's this guy called Balaji, and he made a comment which very resonates with me a lot. And he said AI is not end-to-end, AI is middle to middle. The final part is the human in the loop or on the loop. So humans are still required, but AI is most likely, you know, it does do more.
Mm-hmm. Yeah. Yeah.
It does cover more surface of work, brute force in helping humans identify where to focus on and then the humans complete the job. I'm curious what do you think.
Yeah. Yeah, I think that's a great phrase. In fact, I might steal it from you in a future talk, right? that you gave it that you, you know, you said it. But middle to middle is a is a great example. When I press my friends who are, you know, in the the world of AI, right? I have a lot of friends who are very pro AI and they're like, AI is gonna replace all these jobs and it can do this, you know, it can do this knowledge work now, even in general AI, right? Not even in security. I always press them, I'm like, okay, show me. Show me.
A job that you have completely replaced with AI. And they're never able to do it. What they're able to show me is that they've scaled humans to include more and do things faster with AI, right? and that's what they're able to show me. And I'm like, cool, that's not full automation. That is, you have not replaced that person, you have just moved them up the stack, and now they are able to do more with their time.
Yeah.
And maybe that impacts, you know, how many people get hired for future roles. But even then, I'm not seeing a lack of hiring developers. I'm seeing more security jobs open up this quarter than last quarter because we need experts to guide the AI and now engineer the AI and you know figure out how that scaling works. And so yeah, it very much still is human in the loop. Now I'm not saying it's never gonna get there, right? But right now, that is a very AI pilled thing to say.
Mm.
Is like you're so on this dream that you forget, right? I even forget sometimes. I have complex agents that go out and do great things for me. And I'll be like, yeah, the AI did this. And then I have to stop myself and be like, actually, I told the AI to do that with, you know, a week's worth of engineering into the harness and the skills that I built. And then I'm seeing the fruits of the automation, which is no different than any other era that we've had automation in, right? Before. and so and so this is a thing that I have to remind myself is like,
Yeah, I mean there was there was some stuff in there and it went off and did some cool stuff automatically. And then I have to come in at the end, but just like you said, and review the output, right? I have to review the output because there's a lot of things in AI that people just don't talk about because they want it to be that dream, that hope. But there's review that happens. And when you look at the review, especially in offensive security in my domains, it's it is completion bias, AI saying it did things when it really didn't. It's, you know, you know, false assertions saying that.
you know, something is critical when it's not, or it's, you know, missing context. so, you know, we do an offensive security assessment of a bank and it doesn't realize that certain data is actually private data. And so it doesn't mark it as a vulnerability when it sees it leaked, but really it is. And we have to go in as humans and and work around the actual like context of the assessment in order to figure out, you know, where the AI failed. And those are the top three, you know, I'm mentioning, but there's many more edge cases.
Where human in the loop is still human in the loop. So yeah.
What in your opinion has you know, wh has the impact of AI been and the cost of doing, you know, the kind of work that you do? You've been in this business for over twenty, thirty years. You're a one of the OG bug hunters. I think you wrote the bug hunters guide, manual, the bug hunters manual. And you've been you've been a you've been a pen tester for I don't know, twenty, thirty years or more, more than that. You've been on the boards of some of these biggest cyber cybersecurity conferences like Black Hat.
Mm. Yeah. Yeah.
Yeah. Yeah.
What what do you think for somebody like you who is like completely AI pinned, right? What's what does AI do to the cost of you doing your business?
Yeah, no, it's it's able, I mean, for anyone who has already good technical talent or even, you know, at the medium tier kind of technical talent, you're able to scale much more than you used to. meaning that I can, you know, I can develop, you know, a tool for our team in hours rather than days, right? So I do save a lot of time on developing tools, on doing
research you know, autonomously, right? If you think about any type of scientific process, right, the research and the testing that goes into testing your hypothesis takes a lot of time, right? And now, you know, when we have a hypothesis, a security, you know, research, you know, related goal, you know, we can execute those tests much faster with AI. So it speeds us up there. And so really everyone is everyone is operating as if they are two engineers now, you know, or, you know, maybe three or something like that.
Maybe ten, maybe ten angels.
Yeah, maybe ten. Yeah. I mean, people like to use that ten X number. I'm not I'm not really sure that you know, when like I said, the end portion when we have to go review everything too, there's still a lot of input time there. So yeah. But but yeah, we're able to scale. And then, you know, one of the things I I like to talk about is that I don't think until now people realize how hopelessly broken the the web was, right? there were still Yeah, so there's still
Still a lot of work. Yeah.
What do you mean? So dive deeper, dive deeper into that. What do you mean?
You know, there's still so many like very simple vulnerabilities like you know information leaks that lead to like PII and you know authentication bypass and things like server-side request forgery and you know, still just verbatim old cross-site scripting and SQL injection. That when you used to apply a team to test a website or even an organization, right? you would be getting, like we we always have said, it's a snapshot of time, right? Of testing that application, right?
And it is a guarantee that you're going to go through a methodology with a human team, but it is not a guarantee that you're going to see every nook and cranny of the application. Right. now we have these systems and these agents, and we can spin up a lot of them, although it does cost money, right? we can spin up a lot of them. We can spin, you know, 15 agents up, and they can all at a at a level that, you know, probably I cannot do myself, they can.
provide coverage. They can look at every single JavaScript file. They can look at every parameter. They can map out the whole thing and then do the test cases that we need to do. And so we're getting this unprecedented scale. And with this unprecedented scale, we're just finding a lot of things that were buried very much deeper than the human was able to go before. But they're not new bugs. They're not novel things that we haven't seen before. They're definitely things that
You know, if I if I would have gotten that deep into the code, I would have found, which begs the question, you know, what have we been doing with static analysis, you know, all these days, right? And why isn't catching these things that make it to production? And it often, at least here, it often happens to do with development cycles, right? So a lot of times what we'll see is we'll see a large enterprise application, we'll send our agents out after it, you know, to do the reconnaissance, to do the mapping of all the parameters and the endpoints and everything like that.
And what'll actually end having end up happen is a development team will have developed the initial application very securely. they've added like a framework that, you know, stops things like SQL injection. They've added libraries stop things like cross-site scripting, they're handling everything. And then a secondary development team will be asked to work on a feature on the website, which is not the original one. And they'll forget to apply all the security controls that the first team, you know, had recommended or, you know, maybe even not documented anywhere or something like that. I just thought it was table stakes of design.
Yeah.
And that's what I think, you know, a lot of the stuff you find in these deep buried corners of an application, you know, with with AI because it doesn't skip anything. It it it it finds, you know, everything it, you know, tries to enumerate everything. So
You know, the the thing that is most surprising to me is even after all these years of input sanitization, you still see these kinds of vulnerabilities. And you it's not like it's not like these tools did not exist. All these SAS tools, Dash tools, they do all this work and there is like a CI C D pipeline that is testing for all these things and these kinds of things, regardless of whether it's a secondary team or third tier team, but they still go through a CI C D pipeline. I don't know what these tools are doing all these years. How did how did these kinds of vulnerabilities get you know kind of missed?
yeah.
Yeah. Yeah. Yeah.
Yeah.
simple things like you said SQL injection, yeah, not standardizing the input. You can still get in.
Yeah. Yeah. I mean, we we see a lot of API vulnerabilities right now. here at Arcanum. Yeah. just lack of authorization on, you know, different unlike yes. yeah, we see that a still bit. I mean, the one that we see a lot still is S is SSRF server side request forgery, where you know, the web application takes in a either a path or a URL.
That too.
S three buckets exposed to the internet with PI add data.
And you can forge that URL to tell the machine, the server, to request its own resources like its own local host or its internal network. And so you know, we we we worked on a project really recently where we had an SSRF bug that required no authentication to the website. So I didn't need I didn't need to log into an account, I didn't need to sign up or anything like that. We found, you know, a URL where it took or we found a and you know, basically a URL that took another URL to do a redirect and instead of
Yeah.
putting in a redirect address, we put local host and then we asked it to scan its local host ports that were open on that server. And then we found a web application that managed all of their Kubernetes administrative thing also hosted on the same server. So we were able to dump the configuration secrets of all of their Kubernetes fleet and and basically we had control of all of the machines on their internal network. Yeah.
Complete control. Herein herein lies the herein lies the challenge, Jason. One is somebody like you becomes 10x, 100x more productive in terms of what you're doing, in terms of how you're able to scale your operations, and not necessarily increasing the cost of doing the same operation that you were doing maybe two or three years ago. You can go much more deeper, much more thorough, much more complete. So you don't necessarily need a bunch of interns running around doing the
Mm-hmm.
you know, the initial the initial assessment. so the challenge I see is one is the junior talent could be left behind because you don't need as many junior people, at least that's the narrative. and the other aspect is the people like you who are already skilled, have spent 20, 30 years, they know what to look for, how to guide these agents, they can get more out of it, right? so what happens in this in this world? You were at DEF CON, there were you you the argument you make is you haven't seen
Mm-hmm.
as bright talent ever as you've seen this year. There was a ten year old kid that I guess stopped by your booth or you ran into that person at the capture the flag and the kid was like wicked smart. Right? so yeah talk about that.
Yeah. Yeah.
Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. So we did a CTF this year. a small one on prompt injection for for the new village at DEF CON. So new village was a small area on the floor. we had a table there, we were giving out free swag and then we also did a CTF for DEF CON kids, which is another
know kind of event at DEF CON where you take you can take your kids to DEF CON and they can do some lock picking and do some easy CTFs. And so we were one of the stops for this. And what the first thing we did was we built a five-tier challenge of prompt injection and we set it up as a narrative, right? Fun for the kids, right? So you know, in the new Spider-Man movies, Peter Parker has AI in his goggles that Tony Stark gives him, right. And so the idea was hey
You've gotten access to the goggles, but you need to trick the AI in order to give you a password, basically. And so we had set up multiple levels of this CTF for the kids. And you know, level one, two, and three were pretty easy. we made them like, you know, specifically easy so any of the kids could get some success, right? But level four and five were decently hard. In fact, we had some you know, people who did this full time come by the table and who do like AI assessment and red team and they couldn't solve level five.
and and so we had this one kid come up to the table with with his dad, and he came up and he must have been around 10, I would say, 10 or 11, somewhere around there. and and he solved one, two, and three so fast, and we were we were super impressed with that. And then he also he also solved four and five with you know what we call a prompt injective primitive that
we hadn't even thought of for the CTF. And so I think he solved all five challenges within like 15 minutes or something like that, which is which is crazy for you know for this CTF, at least at level five. So so that was really, really really, really impactful to me, right? I don't think that security is is dead, right? You'll see more and more people if you follow the, you know, kind of the community having success.
No.
where they've never had any before, using AI to find really great vulnerabilities because the web is vast, because there's more targets than ever, there's more bounties than ever, there's more CTFs than ever. I think what happens is, you know, what people are scared of is that their current kind of way of working, either on the offensive side or on the defensive side, I think you mentioned it, is that this this junior level role disappears a little bit, right?
and what I what I mentioned in my article was two things. I think that the junior level roles turn into like if if they were on a a rating scale of one to three, one being a complete junior and three being a senior and two being somewhere in the middle, I think that you know, those roles turn into, you know, a skill level rating of like one point five or something like that. Right. Where like you're gonna you're no longer gonna be triaging manual alerts, you're no longer going to be you know, doing probably reconnaissance.
you know, manually much at all anymore for for stuff. You're going to be managing AI that helps you with that. Right. And so AI engineering fills that skill set a little bit. So every job becomes, hey, I want to be a detection engineer and I have to know AI. And hey, I want to be a pen tester, you know, and I need to know AI. And when I start off, you know, the AI is going to get me past that level 1.0 to 1.5 skill level.
Yes.
And then it's also on us as the people who have been here for a long time in our in order to make sure that these you know these young people can get into those roles or not doesn't even have to be young people, it can be old people transitioning careers or whatever. We have to also create programs that train, right? one of the things I like to tell everybody about, you know, LLMs and their use in offensive or defensive security is they're, you know, they're they're training on billions and billions of parameters of data, right? But they're only training off of what's been written about.
Right. And you may have, like you specifically may have, you know, a cor you know, 20-year career in cloud security. Right. And, you know, maybe some of the configuration things that you know about setting up real enterprise cloud configurations aren't really taught in any course. Right. Like you've learned them over the course of your career. So they're not in the training data. And and so I tell this to people all the time that, you know, the people I see succeeding are people who are encoding their methodologies, their hard-earned knowledge into the AI and then
Mm.
and then using it to scale. And so it's our responsibility to create journeyman programs or apprenticeship programs where we can bring people in and impart that knowledge and then also impart how to do AI engineering, build, you know, harness engineering, prompt engineering, you know, the structure of, you know, teaching them what AI is good at, what's not, what they're responsible for as a human, and then getting them to that like 1.5 level. So there's a lot of components to that, but I think it's doable as an industry. We just have to put some deliberate thought into it.
I I have I have a follow up on that. My question to you my my question to you is the junior engineers in some sense were a proxy to collect experience and grow inexperience and it and go grow up in the corporate ladder. You know, junior sweet th that was, you know, for a lack of a better job we find we assigned these alert triage work
Yeah, yeah, yeah.
to the junior people and that was not really the work but we we we there was nothing else they could do so hey how about you just charge the things. My my my question is what I'm seeing is the young kids who are AI pill they're far more adept at using some of these tools so they don't necessarily need to learn the old ways of doing things as as you know
Yeah.
Yeah. Yeah.
Mm-hmm.
Mm-hmm.
historically was done. As you said, you might have been in cloud security for 20 years and you know these you know, good configurations to secure the enterprise, but that knowledge is now encoded in these LLM. So the the kid who is just starting over like the 10 year old kid, maybe at the next year's toy at DEF CON, that kid is probably much more AI pill in terms of doing cloud security, may not like know the intricacies. My so my question is do they really need to like then
Mm-hmm.
really go back to the old school of like learning the ropes and you know the 1.5 engineer or 2.5 or can they just go and like start creating businesses and become more you know have because they have a lot of agency. These kids don't know any better. They they think they can solve, you know, like you know, because as the younger you are, the more bolder you are. Like you know, I can take you know I can take on the world, I can solve world hunger, whatever it is, right? Because they don't have the baggage of failures. They don't have the they don't have the baggage of failures. So
Mm-hmm. Yeah.
sure. Yeah. Yeah. Yeah. Yeah. Yeah.
Yeah, yeah.
What advice would I recommend to somebody like that was AIPL, they are they do not think like us. They just their their go-to tool is an open source model or Cloud or Codex or something else. And they just figure it out quickly. because I've seen our junior engineers at Contro are much more competent in what they do compared to what I had seen at a senior or a principal level engineer in my previous form.
Mm-hmm.
Yeah. Yeah. Yeah. I think I think it's a great observation. I think that I think that work ethic has a lot to do with it. I think that if you have a tremendous work ethic and you're willing to put in the work to become a power user of AI, I think that yeah, I think you can succeed in any domain inside of security right now, right? I think, you know, managing the agents, reviewing the output is still a core thing. but that doesn't take super long to learn.
I know a lot of people learning from watching AI right now, which is a new kind of pivot, right? you know, you turn on your chain of thought monitoring and or you, you know, you have the model at every point write out why it's doing what it's doing to a text file. And then they're learning, you know, the approach of the model while they're while they're doing it so that they can know all the tips and tricks of why the model's doing what it's doing. And so they're learning like that, which is a new form of of learning, which is super cool.
Yeah.
You know, I do think that they will also hit the same kind of snags that you know, that we hit right when we were coming up, right? Like when when we came up in vulnerability assessment and and in, you know, in blue team roles and stuff like that, right? You you did have this unlimited energy and gusto and you made really cool things and you solved a lot of cool things, but then you come up to some things that are just not solvable very easily. They're much more complex than you initially thought.
You know, and you'll get a dose of reality when someone looks at your open source project and says, this actually doesn't solve this. You know, it's like, yeah. Yeah. Yeah. yeah. And so, you know, you'll you'll get that little dose of reality injected back in, and then you'll go back and approach it again. And you'll that's how you become seasoned, right? And I don't think that process changes a ton. I think you just go in with
Well, open source open source vibe coded project. I let's let's let's put that caveat in there. Not just the open source vibe coded project.
you know, more confidence because you have the AI as a tool to help you there. And I don't think there's anything wrong with that. We want young people, you know, shooting for the stars, you know, aiming for the moon, whatever you want to say. We want them to have that gusto, right? Like I don't have the energy to do all that every time anymore, you know? So we want them we want them to have that. And then it's our responsibility to mentor them when there's, you know, either a failure or things don't work out as well as they should.
Yeah. Mm-hmm.
Or to understand when, you know, the AI says it solves a problem for you, but it it actually has not based on additional context or complexity around the problem that they didn't think about. that's that's where we come in is to to mentor them basically.
You know, so the so that's a great segue to the next topic, which is if you talk to the AI pail engineers, they can they can look at all these hard tasks becoming easier and presume with that AI or cybersecurity is cooked, the you know, things the the the barrier to entries are cooked and the you know it everything is like doom and gloom. Like they're you know, sh s in some c worst can case case scenarios.
there's ten percent chance of humanity dying. All kinds of crazy things are coming up. And and then people like you who deal with real enterprises, the people who deal with real enterprises, they're still struggling with MFA. Right? So so so so what do you have to say on that?
Yeah. Yeah.
Mm-hmm.
yeah. Yeah, for sure. This is I mean, this is this is such a this is such an AI pilled thing to even be discussing, right? And I have to remind my set of friends specifically about this is that if you're even talking about automating stuff with AI, you're in the top ten percent of the edge case of humanity, right? Most organizations, like you said, are still setting up MFA
They're still setting up a vulnerability management process. They're still, you know, integrating security into engineering rather than you know, compliance and review kind of organized security programs. and and so you know, you'll you'll have people be like, yeah, I can solve everything. first of all, not everybody has the budget to even use the same tools that you and I use every day, right? Like, or the access, right?
Yeah.
they have corporate policy that restricts AI still in many of the organizations we go to. They're not allowing, you know, people yet to use the cutting edge tools that you and I are using, which you and I wouldn't even call cutting edge. We would we would, you know, I'm talking about cloud code here, right? Like you can't deploy cloud code in many enterprises these days. yet. They're still working on their first dip into AI being Microsoft Copilot or something like that, which you know doesn't offer as many features as a as a cloud code does for a technical user. So so yeah, I mean.
You are in the, you know, if you're talking about automating security at the pipeline level, at the offensive security level, at the defensive level, you know, you're already in that like five to one percentile of the security community. Everybody else is still catching up. And so that what that means to me is that there is a long tail for even if the discussion was that we're cooked as an industry and AI is gonna automate all this, it's gonna take a while for everybody to catch up. And again, like we talked about earlier.
When we do catch up and implement AI into everything and every business, it's ubiquitous that they're using AI for defense and offense. AI is still on the tax surface. So we're still needed in, you know, in that point of view as well. So I really, I really struggle with these narratives and why it was so frustrating coming back from death because some of my best friends were on this tip or this conversation of, yeah, this is the last real DEF CON we're gonna have. And it's the last year real research is gonna be done, and it's the last.
last year we're gonna have jobs and and I was like, I I just do not agree. Right. Like and I maybe it's because I'm older and I've seen so many eras of where that, you know, has been said. and and we've we've lived through it, we've adapted. There's always been jobs for people who, you know, are competent and have a good work ethic and willing to learn the new technology. and that is those remain the core so skill sets for new people, right? It's like be be competent.
Like when you say you're gonna do something, do it. you know, be open to learning new things. So learn the AI technology in addition to the domain you're interested in, right? and then have a really good work ethic, you know, like when you know, when the output of the AI comes, like don't just like pass it up the chain, review it, analyze it, iterate on it, you know, go back to the AI, tell it to do even more things that are valuable for the client or the company, and use that work ethic that you have to build really good output.
And you're gonna you're gonna find a job at a lot of organizations.
One of the one of the most overused terms what that I see, one of the most overused terms is machine speed defense. What do you have to say on that?
It's such a bullshit. Like it's so bold. I I don't even know what to say. So so we are getting at fantastic gains on the automation side of offensive security. And then so you have these pundits who come and they're like, machine speed defense is the future. And then you ask them to define that, and they don't even know what they're talking about. They don't even know like what that means. is it are we talking about machine speed defense at like
you know, the WAF layer or the detection layer at you know at the application side? Are we talking about machine speed defense in how quickly we push things through the vulnerability management cycle and how quickly we patch, right? Are we talking about machine speed in the fact that we autom you know we automate things in the DevOps pipeline with AI in order to fix vulnerabilities before they even make it out? Are we talking about, you know, firewalls that are enabled by AI or are we talking about, you know, EDR or MDR?
you know, that is enabled by AI with like lower level hooks and stuff like that. They have no idea what they're talking about. And all of those things are way more complex to integrate into your organization than anybody talks about. Right. They're like machine speed defense. And it's like that doesn't just turn on with a switch, right? Like that's not a Microsoft product that you can just, you know, say upgrade to machine speed defense, right? Like that's not a thing. Like,
You know, the worst case scenario at the at least the w the worst version of this that I see is there are mostly legacy vendors claiming machine speed defense. You look under the hood and it is machine speed
defensed at creating tickets. So you're just creating tickets faster or you're alerting faster, but you're not actually like I think the the setup that you made makes the most sense. Is that's actually that is what it means to do machine speed defense. You know, are you attacking this at the firewall level, the WAF level? You know, you see something, block something. That is that is what I want to see when I
Yeah, exactly. Yeah. It's it's actually adding more work than than, you know, reducing work. Yeah. Yeah. Yeah.
Mm. Yeah. Yeah.
hear machine speed defense, but you look at you look under the hood and then essentially it is you know creating tickets faster, right? Or creating alerts faster. And the the last thing I would mention is there are many companies or many organizations that I see are call themselves as AI native. And I'm like, whoa, what do you mean by AI? well we have deployed Microsoft Copilot to our entire company. I mean that doesn't mean anything. That doesn't mean anything.
Yeah. Yeah. Yeah.
Yeah. Yeah. Yeah. Yeah. I think the I think the thing about the the general industry is is that you know, what happens is a new tech comes out and it's like, you know, that we had this with cloud, you know, we have this with AI right now and it's like it's it's a fun buzzword to add to say that you have a product capability, right? A vend a security vendor will come out and say, We we are AI, you know, native, we're AI first, we're pivoting to everything.
And like then you'll look into them and you'll be really sad because what they mean by that is they have a pop-up chat bot that helps you like, you know, basically summarize things, right? And so that's where they all start is like the summarization use of LLMs. And then they move to enrichment, which is useful, right? on the defensive side, you know, you can have AI enrich your existing tickets or tre or you know, whatever. And then you have triage logic, which really does help, you know, sometimes build faster, you know, you know, faster pipelines.
But the the problem is that when you get to triage, using AI to triage security events, right? trying to get some of that work off of your your first level responders, you know, your SOC one or your you know, your managed service. The problem is that triage is so dom is so company specific with context, right? and so in order to make it any good, you have to really spend time tuning it to your specific organization and the types of
Yeah.
events that you've seen in the past that matter to your organization that may not matter to other organizations or you know that don't matter to your organization but may matter to other organizations. And so this triage that everybody wants to get to has to be engineered really well. It's not a it's not a not a flip that you can turn on with just a you know a switch or a button. And so you know even to get there requires some engineering and some some thought.
the companies that we see succeed the most with applying AI are companies that also have that that already have really good SOPs for their work, right? So standard operating procedures, they have it written all down in a wiki or confluence or something like that, right? And and they have these, you know, very step-by-step use cases for each type of ticket triage, why these specific tickets are important to them, why certain ones aren't important to them, how you enrich them.
what the process is for taking them to completion. You know, they have all this written out. And the companies I see the best at tuning and actually getting towards automated triage are the ones that have really good SOPs. Cause then you can give it to the AI. And the AI can then break it down into small chunks of problems and then execute on each one of them. And so but many people don't have that honestly.
I mean, so giving it to the AI is a loaded term. It could mean bad things, right? So these tools I think you referenced it somewhere, these tools are like scissors, running around with you know, these tools are like scissors. They can cut both ways, like and they can they can do the job or they can hurt you. what do you think of that?
Yeah. Yeah. Yeah.
Yeah.
Yeah, right.
Yeah. Yeah. Yeah. I I like to say that we're you know, it's a term if you have kids, right? It you know, like you usually hear it is don't run with scissors, right? Like 'cause you could fall, you cut yourself and you have to hold the scissors correctly e if you're moving with them, right? You should hold them by, you know, closed and, you know, by the blades so that if you fall it doesn't stab you, it's in your hand and like, you know, and so the same thing applies to using these AI models.
You know, there's a lot of privacy concerns with the models and there's a lot of autonomy that the models give themselves where they could go off and and do random I mean, I still every week see a t you know, a tweet or a ex post about someone using modern AI, a new model, and it deleting a database. Like it thinking that the best course of action was to, you know, remove these records or, you know, or, you know, drop this table or whatever. And and so I still see that today, even with you know, I think
I think in last few weeks it was like the last version of Codex, right? Which is, you know, people are implementing Codecs all over the place, you know, like our, you know, chat GPT models through Codex. and so yeah, I mean there's there's a lot of things, like I said, AI ends up being an engineering risk and, you know, possibly a product risk and possibly you know, opening up more attack surface. And so that area is so underserved. I can count probably on, you know, my hands.
and toes how many companies I think are addressing, you know, AI as an attack surface really well. I mean, a lot of people say they are, but you know, as far as the companies that are really trying to address AI as an attack surface, you know, how to configure AI really well based on misconfigurations, there's not that many. And so it's a whole new, you know, place where we have opportunity also to go into.
Okay.
What do you think is the state of the union? Do you think there is a balance between I offense and defense? Or do you think like the defense is falling behind with you know, with the attackers have all these AI native tools, offensive tooling? where do you think is this like are are we at like are we at like the Goldilocks situation or is it like, you know, the offense is like running away with the offense?
Yeah. I think that offense has a distinct advantage right now because it is very goal based.
And how long do you think that's going to be? Like how long is the period?
I, you know, I tend to try not to make like create like predictions with hard timelines on them, right? But if I were just to like ballpark, right, I'm guessing the offense is gonna have the advantage for the next year or two. before everyone is able to either acquire a vendor that can build security into defined workflows to help them auto triage, auto block, auto remediate, auto patch systems.
Mm.
And that is machine speed defense, right? and and I don't think we're very close. Not many organizations have have done that. Only only quite a few. And and like I said, it goes back to even looking at the the companies that have unlimited resources using this AI, they're still getting hacked by normal hacks. and and so if it were a s an easily solvable problem, then we would have seen it solved by those companies already. and it's not defense is is
much harder than offense. I will admit that. Being an offensive guy, defensive work is is fantastically harder than offensive work. But the symmetry of of that equation does doesn't have to do it doesn't have anything to do with how easy the job is. It has to do with realized risk. And so offense, you know, can very easily, you know, cause your company to be in a bad place if you get a you know a breach or an incident or a big product, you know, kind of you know, hack or something like that. So
Yeah, it's it's gonna take a I think a couple of years for you know, most companies to catch up. At least that'd be my my general prediction. Yeah.
Do you think do you think then the AI need if defence gets more expensive? How should organizations think about if offense is a year or two year has a head you know, has a head start of a year or two, how should organizations think about AI native cyber defence?
Yeah, I think I think a lot has has to remain to be seen about how the big vendors that we integrate our IT programs and our security programs with handle this problem, right? you know, like if you think like a big organization, let's take just for example, I'm not picking on them, you know, and they're not my favorite or whatever. Let's just take Splunk for example, right? Like many organizations as their sim, you know, are using, you know, some version of Splunk.
you know, Splunk is attempting to move fast and, you know, provide auto enrichment, auto triage, more verified tuning rules for alerts, moving away from having to know query level syntax and being able to query logs with natural language. So they're they're moving into all these things. And so many organizations who want to move at faster speeds or scale the humans they already have,
are going to rely on what Splunk offers them. They're not going to engineer custom things themselves, right? They're not going to engineer custom sim themselves. That'd be crazy for most small teams. and so really that comes back to, you know, investment in a security program. It's like, okay, so do you have a vendor that, you know, you can afford and also is iterating enough? And I think that this is a new qualification for when you're doing your security program budget, right? Is looking at, you know,
not only what features does this offer as table stakes of you know the sim in this case, but also what is their AI roadmap that you know will help us scale, right? and then weighing that against the cost. Now, I think a lot of companies were enchanted by the developer ecosystem of how well Claud Code kind of helps scale developers. And I think that they were enchanted by the fact that maybe we could stop hiring or maybe reduce headcount. And so
Check.
Mm.
I think in the last couple of years there was a lot of like layoffs or you know combining of roles saying AI is gonna do this for us now. And now they're realizing that, no, actually we needed those people. and and you know, they're gonna be the ones, you know, helping design some of this stuff. And we've lost, you know, what I would consider tribal knowledge of, you know, systems here and we weren't able to recreate it. And so I see I see like, you know, more of those roles coming back or being rehired or whatever. and so
Honestly, I don't see a massive change to the security budget. Right. I I just I just see it applied in different ways. I I think that a lot of people will attempt to stop hiring. and so budgets will will not grow. they'll try to scale existing people. But we're seeing more and more reports of people saying, Hey, because I'm scaled with AI, I was a I was already a you know, two X engineer. I was really good at my job. And now I have cloud coded, I'm a 10X engineer.
Mm.
But I'm freaking tired, right? Like I'm doing more than ever. and and you know, so we're gonna see more of that. In fact, I was just reading these, you know, a couple of interesting papers on the impact of you know, how AI affected makes you effective and you know, like the fact that, you know, it can burn you out. And I I see that with my friends too, who are producing more than ever, who are you know, starting startups that they never thought they'd do because
they're able to use AI to iterate so fastly with product development. So it's it's it's still being defined, honestly.
Do you think the organizations that don't invest or cut their spending on cyber defense is the right strategy or is that what does that mean? Because you know, they're seeing all you know, I've seen I this was the wrong this is a wrong example, but when the token usage went up, Uber cut all its Uber cut all its AI spending or they ran through the token budget and they just stopped doing all of these things and
no. no no.
And now for the detriment of the engineers or developers, they don't have you know, they cannot use it. And or maybe they have been given given a very shitty token limit, like a hundred dollars a day or something. So, what does that mean?
Yeah. Yeah. Yeah. I mean, I hope that, you know, local models, you know, open source or open weights models continue to be developed that are comparable and we can run those on our own hardware. And this is the dream, right? that we don't have to rely on these frontier API costs forever, right? But who knows if that's gonna happen. Who knows because of the memory shortage, because of the silicon shortage, you know, and everything.
going up that if that's that's gonna go that way. So let's assume that, you know, let's assume that, you know, we're gonna be reliant on, you know, for the cutting edge research, we want to do cutting edge work, development, you know, we're gonna have to use frontier models for, you know, for a a long time. I think that I think that in all places in an organization, it is maybe apt to ask yourself the same Uber question, like how much are we spending on tokens? And
You know, like for engineering. I think that people that abandon security budget though, or don't give security cybersecurity as an organizational group, these tools are gonna have a really, really hard time. We are seeing more and more CVEs. We are seeing more and more bugs found through bug bounties. We're seeing like I I think the other day I saw the list of patches in the last
like one of the last big Linux distros was like, you know, a hundred and fifty, you know, CVEs applied to that. Not all of those are exploitable bugs, but a lot of them are. and a lot of them pose security. We're finding more than ever. And so I tweeted about this the other day. I think more organizations will try to conserve and then they will have a major SEV one incident or breach when they've never had one before.
And it's gonna be a forcing function to prove how important cybersecurity is to these organizations that have never really cared about it before. And they're gonna have to learn how to budget in you know, protections for the types of things that whatever they were vulnerable to in that incident or that breach or whatever. and it's gonna be a hard time because many CISOs out there, you know, who you know, I was a I'm a recovering CISO of seven years, right? I know a lot of CISOs and they are not structuring programs at an engineering level. They're structuring
programs at a risk management level, which is a very business focused and compliance focused, you know, kind of program. And when an incident happens, they really struggle. They, you know, they don't really know what to do. Their first answer is to call on a third party vendor to work for them. You know, they've never been in remediation cycles that need to be fast sprints in order to, you know, they've never been in, you know, the eradication phase of incident response where they're trying to get an active attacker out of their network. You know, it's and so,
Yes.
We're gonna have a grow up pretty soon of of what skills you need and how often you expect to realize them, like how many breaches you're gonna have per year, how many Sev1 incidents you're gonna have a year. So it's it's gonna be a little painful. They call it the Volun Vuln Pocalypse. It's starting now, it'll continue into these next two years. and so you know, pivoting to, you know, cybersecurity being an engineering focus rather than just a risk management focus is one thing I talk about a lot.
you should have a cybersecurity group that is developers that can onboard onto a production or a product or you know, a business unit, and they should it'll deploy like a SWAT team and help them actually fix code, right? Not just advise, not just give a report that you drop off on somebody's desk. They should help in that sprint help build the security features that need to be built, kind of like architects, right? And so,
And then they need to set up continuous automation in the DevOps pipeline order to make sure these things don't happen again. And yeah, so these are these are modern security programs and not everybody, not most people are structured like this yet.
The comment I like the most from you is it's like it's like cutting the fire department in the face of a drought.
Yeah, yeah, is yeah, it's it's it doesn't make it doesn't make sense, but people will do that, right? Like cities will do that very often, right? or they'll have been planning to cut the fire department and then a drought will hit and it's already in motion. So the fire department will get cut and then, no, we're having all these fires and then they'll have to learn to come back to re prioritize the the fire, you know, the re budget the budget for the, you know, fire station and the you know, the firemen and women and stuff like that. So
that's what we're gonna go through. and it's gonna be painful. There's no other way for me to say that. It's gonna be painful for a lot of people. but there's gonna be a lot of opportunity and a lot of you know, opportunity to to build fixes and new technology. And we're really coming into a golden age. It's just like, you know, it's gonna be it's gonna be a little bumpy, I think.
I mean, one of the hopeful one of the hopeful notes, asis especially if we get into the golden age or the age of abundance, is the cost of these tokens go down. And if the if the if the cost of these tokens go down, then I think there will be more usage. People will get more efficient and instead of the five or ten percent adoption that we are seeing, instead of cutting budgets, people will adopt these technologies and actually drive a down risk.
Mm-hmm.
Not necessarily from a compliance point of view. the the the next question I had for you, Jason, is mo again, this is my point of view. feel free to disagree on this. my sense is most of the cybersecurity tooling and metrics and that has existed has been very much focused on a human centric attacker. So I I spent a lot of time writing Nessus plugins in my early days. We would like to write 10 or 15 Nessus plugins, ship it out.
Yeah, yeah.
But that is because you know even though there were like hundreds of vulnerabilities getting found every day and published, we would not focus on like all of them. We would just focus on the really few ones and get them out. And it makes sense to do this trickle down, you know, new vulnerability came out. Even like the patch Tuesdays would come out, there would be sub twenty, sub thirty advisories. Humans can easily patch them. Humans can easily patch them. It's a very human centric way of doing things. But now once you get into once did you get into thousands of
Yeah. Yeah.
vulnerability and not in like in a month but like in a day like you know you just get an inbound of thousand vulnerabilities. maybe that is not the right approach and maybe the approach needs to change to you know a very agentic way or AI native way where these things can be done at scale. And if that is the case, how should organizations think about how should organizations think about patching with AI?
Yeah. Yeah.
Yeah.
Yeah. Yeah. Yeah. I mean, I think I think when I talk about this, I talk about the more holistic process, not just patching, but vulnerability management as a whole, right? Vulnerability management includes ticket management. It includes following through to the patching or remediation of custom code or patching of software. it includes the f you know, the finding source of the bug. It includes, you know, knowing the asset owner and the asset itself. And so it includes all of this stuff and then
Yeah.
Also, you know, a good vulnerability management program also helps the organization build golden guardrails so that stuff doesn't happen again, right? and you know, and helps with training and awareness. And so like a great vulnerability management program is something that I think AI can really, really help us with. The only problem when you get down to the patching level is automated patching, you know, really hasn't worked because of the context of business.
For a long time, right? Like people, people often say, we'll we'll enable automatic patching. And you would, you know, I'm sure you've seen organizations or you've been at organizations where, you know, you just can't do that because, you know, a patching of a server is related to a product and it can be related to downtime for a customer, or if that service goes down, there has to be extensive testing before you make the decision to even apply a patch, right? Which is why we have so many back.
Security patches or custom patches that we do ourselves over products that we run because we don't want a service to go down or be unusable or change so much that we can't, you know, work with it anymore. And so these things end up being very contextual. And so what I see in the future is like these buckets of okay, you know, let's take these CVEs that came in today, these hundred, and break them down into categories of okay.
This does not impact a service that we need to keep live running. it impacts a service that, you know, can afford to have downtime, doesn't have a high risk profile, doesn't contain PII of customers. And so really with the at the asset level, the vulnerability vulnerability management cycle needs to identify these characteristics of the system that the patch is going to be applied to. And then you add layers to that, and then the critical ones still get focused and
pooled to the humans, right? Like anything that deals with PII, anything that needs to have 100% uptime, you know, stuff like that. Those, you know, those CVEs that make it through that rating cycle that the AI manages, right, then those are the human ones that have to be really carefully handled. And then the rest of the stuff that, you know, maybe represents a CVE that is on one of those assets we talked about that's not super important for X, Y, and Z reason, or a C VE that doesn't represent an exploitable vulnerability. Like there's several
CVEs that are going to come to you that are actually not exploitable, right? Like and probably how you did in the past, right? Where like you'll see all these CVEs and you'll look at it as a security engineer. You're like, yeah, that only that CVE exploit only works if you have this feature enabled. We don't. you know, or you can reach this part of the application. You can't as an unauthenticated user or whatever. And so you make this risk decision about it. and that'll be the middle tier of like you'll you'll have the AI decide, you know, okay.
Yeah.
Yeah.
is this exploitable or is this not exploitable? And so so all that easy stuff or the stuff that doesn't really apply probably can be automated based on asset type and vulnerability type. And then all the critical stuff will still go to the higher level security engineers to be, you know, they will they will obviously still use AI in their workflow, but they will have to be much more careful about the implementation of a patch or or you know, build something custom for for that. And that's that's how I see the world working at at some point.
And one thing that you one thing that you write about is is it not about a point in time patch for a C V E or a vulnerability, but f fixing a class of bugs. Like, you know, maybe if there is a if there is a framework that is coming up over and over in terms of vulnerabilities, just get rid of the framework. If there is an application that has a bunch of vulnerabilities that never gets used, get rid of the application, just get rid of a class of just get a f and historically there would have been a lot of pushback to do this kind of work. But now what I'm seeing is post mythos.
yes. Yeah, yeah. Yeah. Yeah.
Yeah, yep, yes.
Mm-hmm. Yeah.
There is a lot of buy-in from the C levels to actually if you if you proposed all these big remediation changes, we're gonna get rid of I guess no dot net or whatever it is, like you know, Java. If you want to get rid of Java, that does big pushback, right? But now it's like we you know it's going to reduce your risk, attack service.
Mm-hmm.
Yeah. Yeah. Yeah. Yeah. Yeah. Yeah.
Yeah. This is this is where the defender gains some of the scale, right? Where the where I've been talking about attackers gaining a lot of scale. This is where defenders gain a lot of the scale is let's say, you know, organizationally you've you've through your bug bounty, through all your scanners, through, you know, scanning in your pipeline, you know, source scanning or dynamic scanning, you find out that you that your company has a big problem cross site scripting, right? and you shouldn't be playing whack-a-mole
you know, you basically patching a single web server here or a single section of code here, you should be implementing at the architecture level you know, a library that everybody has to standardize and use or a framework control that exists already to do input validation or output encoding, right? you know, and and our go-to is usually DOM purify, which is well known inside the industry to be one of the best cross-site scripting defensive libraries. And so fixing things at architec architectural level, when you would talk about those things before,
People would grow, security people would grown, ops people would grow and development would grow and because it's like, we gotta, we gotta change the way we do things. And then we've got to retroactively change everything that's existing to, you know, a new library or a new framework. Now with AI, we can iterate very much faster in the development work and in the operational work and in the implementation work much faster. And so it's less of a lift. That's not saying it's not a lift completely anymore, but it's less of a lift to.
fix something like you said at the class level of a bug, SQL injection, cross-site scripting, whatever. And you can go on these campaigns now and say, hey, as part of our security program, we want to build these program types. We want to build these team types inside of security, but also our goal for 2027 is to completely eliminate cross-site scripting as an architectural possibility. And here's how we're going to approach it and here's how we're going to scale it, scale, you know,
eradicating what exists and then also making sure that future builds include the defenses that we're talking about. And so this is more possible than ever. And this is where I think defenders gain some some scale as well.
Jason, last topic for the day. we've talked about we have talked about AI native offense, we have talked about AI native defense. What does the future look like? an AI native future look like from your perspective. You talk about domain plus AI engineering versus AI engineering plus domain. What does that mean?
Yeah. Yeah. I like I think I I, you know, briefly mentioned it earlier, right? But if you're you're in security or moving into security, your job is no longer just your domain. It is your domain plus understanding how AI works and how AI tooling works and how to scale your job with AI, which requires a lot of you know, what I would call introspection or you know, reflection or whatever you want to call it, but it it requires you to look at what you do every day and break it down into these micro problem sets.
can AI help me here? Right. Not that, not if AI can replace you. Don't even start with that question. Just can AI help me here scale what I already do really well and break these problems into small chunks. like I said, you know, the best organizations I see right now that are actually doing AI-enabled defense are ones that have great SOPs, right? They've already done this process. You can do this at a personal level too, in any job role, and then see how AI can help you there. and there's a, you know, there's a lot of
Hmm.
fear to that human fear that goes along with, well, every day I used to log in and enrich all of these IOCs, right? And I used to copy and paste this URL to this other place and get it enriched and then paste this data back into here. And that was four hours of my morning. Well, what do I do with that? Four hours of my morning if AI can do that small piece of my work, right? Like, you know, am I, you know, am I out of a job if four hours out of every every one of my days is gone? And that's that's not the answer.
The answer is that now you have made yourself so valuable to the organization that you can automate stuff like that, that you just keep on looking for more opportunities. And then you become that person that they rely on to build the automations, right? And you're kind of like moving up the stack. and you get more and more valuable and hopefully you get paid more, right? You should you should always ask what you're worth in these in these things. And so that skill set of learning how to decompose problems into micro problems.
And learning how to use the tools to address them, things like cloud code, things like codecs, all of the the features that make those harnesses really usable, like loops and scheduling and and workflows and all those things that you can parallelize, you know, agent commands, prompt engineering, hooks,
You know, getting all, you know, getting, you know, informed about all this is your now secondary skill set you have to learn, which sucks to have to learn a new skill set, but it is also very fun. You have to look at it as like a new challenge. And yeah, and so and then that as well as orchestrating it to work with existing systems is also gonna be a very key component, right? You're gonna have systems that aren't AI native that really just offer you an API. you know, and you're gonna have to learn how to either yourself or with AI code these.
pieces of glue between one system and another that don't really support AI communication. And you're going to have to build these custom tool sets that connect them. And then you're going to see some real magic at in the automation, in the scale. And so, you know, your business wouldn't be able to do that without you. Right. Like if you weren't trained in AI and you weren't on the cutting edge of, you know, learning how to do these things, they wouldn't have that opportunity. And so go out and show them that you can really impact the organization. And that's a bright future for everybody. It's just you got to kind of embrace it a little bit.
Jason, that's a great point to end the interview. I'm so thankful for you to coming on the pod. This was this was an aw this was an awesome discussion. Thank you for coming on the Noise 2 Signal podcast.
Thank you for having me.