Episode 996 ·
How AI Finds Novel Web Exploits with James Kettle of PortSwigger
What if the best way to out-hack AI is to teach it what you know that it doesn't? Today, we're talking to James Kettle, Director of Research at PortSwigger. We explore how he turned a decade of research intuition into an autonomous system that invents new attack techniques, why AI has changed what it takes to do original security research, and what happens when safety guardrails and agent behavior collide.
All of this right here, right now, on the Modern CTO Podcast!
To learn more about PortSwigger, check their website here
About James Kettle
Modern CTO listeners can use code MODERNCTO10 for an extra 10% off at https://nordlayer.com/browser/moderncto/
James Kettle is the Director of Research at PortSwigger, where he spends his days pioneering novel web attack techniques. His recent work includes HTTP desync attacks and automating the hunt for unknown vulnerability classes. What he enjoys most is building those techniques into Burp Scanner and then watching other hackers use them to uncover new vulnerabilities. Three of the ten most popular Burp Suite extensions are his: ActiveScan++, HTTP Request Smuggler, and Backslash Powered Scanner.
Kettle only got into hacking because he grew bored of playing Counter-Strike at university, yet he’s now a well-known figure in the community. He has spoken at many prestigious venues, most recently on HTTP Desync Attacks at DEF CON and Practical Web Cache Poisoning at Black Hat USA. Away from the screen, he’s an avid cyclist, often riding to get away from his computer, which he bought with a bounty payout, paid in Bitcoin.
Transcript
# Episode 996: How AI Finds Novel Web Exploits with James Kettle of PortSwigger
(James Kettle at 00:00:00) Maybe if I take my own personal research methodology that I've developed over many, many years manually and build this into a system, I can make an AI-powered system that can autonomously invent original techniques and use them to hack real websites. But Opus 4.6 massively outperformed everything else at the research cascade. So Opus 4.6 is basically better at hacking than Opus 5.5. And that's excluding issues where you hit guardrails.
(Joel Beasley at 00:00:40) I wanted to give a special thank you to Nord Security for sponsoring today's episode. Nord has been kind enough to give us a discount link for their newest product, NordLayer Browser, that we have put in the show notes below. If you don't know, NordLayer Browser gives IT control over browser sessions so that work data stays protected where the work actually happens. I'm curious. I was reading all about you and your background. You've been a security researcher forever, essentially. What got you into security research originally?
(James Kettle at 00:01:11) I think it was through bug bounty, really, because to be successful at bug bounty, you need to find vulnerabilities that other people haven't found when they've already looked at the target that you're looking at. And I always had the belief that everything can be hacked if you look at it hard enough, if you go deep enough. And that mindset drew me into research.
(Joel Beasley at 00:01:41) Yeah. My first, there was this site called Hack This Site, you know, 20 years ago. Are you familiar with that one?
(James Kettle at 00:01:49) I've definitely heard of that.
(Joel Beasley at 00:01:50) Yep. Yeah. Yeah. And I came across... I'm 38. So I came across it when I was... Had to have been 14, 15, and I started playing with it and just progressively taught you better and better ways to attack it. And it became... It was so interesting. And then I realized that, oh, I could make money doing this. But then I also found that I could build software because you kind of had to learn both. You had to learn how the software had to work to attack it. And then I kind of fell in love with building things, and I said, you know what? I'm just going to go the builder direction. So I actually didn't really do much security. Maybe for a year or two, I was playing with it and having fun, but then I went full builder.
(James Kettle at 00:02:30) Nice. Yeah. My approach to breaking things is very much built on building things to break other things.
(Joel Beasley at 00:02:37) You build the... See, I didn't even think of that. I was just a kid. I didn't think I could build things to break things better. Right? But that's what you've been doing for your whole career, isn't it?
(James Kettle at 00:02:48) Yep. Absolutely.
(Joel Beasley at 00:02:50) Oh, very cool, man. Okay. So I'm curious with this rise of the large language models and the intelligence. What's that doing to the security world?
(James Kettle at 00:03:03) So it's causing a lot of chaos, I would say. It's kind of reshaped vulnerability discovery hugely. And right now, we're going through this really intense period of time where everyone's having to adapt very, very fast, relative to how fast things were changing previously. And we're seeing tons and tons of vulnerabilities that have been there for a very long time now getting flushed out within a very short space of time, which is making for quite an intense experience for everyone.
(Joel Beasley at 00:03:43) So you're in the day when you would have to make your own butter, and now there's butter machines. Right? Like, to... You would have to go and research and find these very specific security vulnerabilities, but now you have tools that can do that for you. Is that an incorrect assumption?
(James Kettle at 00:04:00) More or less. No. That's pretty accurate. So it's like, for example, a few years ago, I would have one finding, and I'd be excited about that for two weeks if it was a technically interesting finding. And then this year, when working on the HTTP Terminator, it was just finding things so fast that I'd be like, "Oh, no. It's another finding when I come into work in the morning," and I'm like, "Oh, no. I have to understand it and work out how to report it." And it's like the whole frame is flipped.
(Joel Beasley at 00:04:36) What's this HTTP Terminator thing?
(James Kettle at 00:04:40) So this is my most recently published piece of research. At the time, about a year ago from now, I saw a lot of people saying AI is great for finding vulnerabilities, but it can't come up with something truly novel. It can't do genuine, original research and invent new ways of hacking things. And so I thought, "Well, you know what? Maybe if I take my own personal research methodology that I've developed over many, many years manually and build this into a system, I can make an AI-powered system that can autonomously invent original techniques and use them to hack real websites."
(Joel Beasley at 00:05:23) And you were successful in this?
(James Kettle at 00:05:26) Yeah. Very much. More than I bargained for, really. It went... It found tons and tons of stuff. It went beyond the... So I built it to find novel request smuggling techniques and do novel research in that area, and it actually ended up breaking out of that area and finding an attack type that has implications significantly beyond that.
(Joel Beasley at 00:05:55) Wow. Is this, not to get too nerdy, is this issues with HTTP or the browsers? Where is the vulnerability actually laying?
(James Kettle at 00:06:06) So in request smuggling, it's predominantly issues between the front-end server, for example, Akamai, and the upstream servers of that, such as maybe the application back-end server.
(Joel Beasley at 00:06:22) Interesting.
(James Kettle at 00:06:23) Generally, that hop is communicated using HTTP 1.1, which is a very, very old text-based protocol that is absolutely horrifically insecure. My previous year's research was called "HTTP/1.1 Must Die," but it hasn't died. So it's still a fantastic way of hacking things.
(Joel Beasley at 00:06:45) Tell me how you really feel about it. So you usually have this... I was... You have so many great articles on your site. That's key part of what you do is you research, you publish this research. It's all on there. The best of hacking for each year. It goes back quite a ways. But it looked like you were on this yearly research cycle. Is that going to change now with the speed at which HTTP... Do you call it HTTP Terminator? Do you just call it Terminator? What do you call it?
(James Kettle at 00:07:17) HTTP Terminator. Yeah.
(Joel Beasley at 00:07:20) Okay. You just say it real fast. HTTP Terminator. Alright.
(James Kettle at 00:07:22) Yeah. I've got a lot of practice in saying that word.
(Joel Beasley at 00:07:26) So is that going to change your yearly research cycle, or will it be the same just filled with more? How's that?
(James Kettle at 00:07:34) It's definitely changed how I approach it. My job for the last 12 years has been to do novel research and present it at Black Hat USA, and that has not changed. But what it takes now to do the kind of research that will get me a presentation at Black Hat USA next August has massively changed because it's become so easy to do research now. If I want to present something in nine months' time, that means I need to do something that nobody else could possibly find themselves in the next nine months. Right? Because it has to be original when I present it. So these days, you could... You can do novel research, if you're lucky, with a single text prompt into ChatGPT. Right? So I've had to think about... Really hard about, where does that not work? What will AI not find you ever? And really explore that space to try and come up with something that I think I can land that nobody else will land in that time frame.
(Joel Beasley at 00:08:50) Right. And also, based off of the experience level of your peers, what's the thing they're not going to know how to use AI to do? You know, because I've found there's such an art to how you interact. Two people with the AI are not equals. Like, if you take a senior engineer with 20 years of software experience and you take a person working at Chili's or a restaurant and you pull them in and you give it to them, they're going to be able to do very different things with that AI. And so you're trying to figure out what is the best version of James. Like, what could I do? What's unique to me that I could do with AI to figure it out? And this year, the HTTP Terminator thing was the thing. You put your recipe. You might as well just call it James AI.
(James Kettle at 00:09:38) Yeah. Pretty much. It's that, when you're trying to do research to get presented in a conference, you're competing with everybody else that might possibly have a research idea. I mean, the way AI finds things, it'll find things you didn't even ask it for. You might be competing with people that aren't even trying to do security research. Right? It might just be like, "By the way, I found this thing in your code" and just lead you down the research itself. So, yeah, that's the challenge is getting outside that and going, "What's something that's maybe so ambitious that no one would try it?" That's what the HTTP Terminator was. And now for me, the follow-up is like, "What's something that everyone thinks is impossible so they wouldn't try it?" That's a good example of something that might be safe, maybe.
(Joel Beasley at 00:10:43) Yeah. And then you just have to trust it and hope that something good comes out so you can present it at the conference.
(James Kettle at 00:10:49) Exactly. It's very much a leap of faith. And in a way, AI has made... Yeah. AI means you can try things out very, very fast. But if anything succeeds quickly, then it's not good to be presenting in nine months' time because the chance of someone else doing it is once again really, really high. So, I mean, kind of... Yeah. Exploring this space of what's genuinely difficult, and maybe people think it's impossible, but it's not quite impossible.
(Joel Beasley at 00:11:21) And you'll figure it out right before submissions for the conference.
(James Kettle at 00:11:26) Well, we'll see what happens. I found some things already. So that's good, but, obviously, I can't talk about those just yet. But, yeah, I've got some optimism.
(Joel Beasley at 00:11:42) They already had the 2026 conference. Right? So you're just preparing for '27 now.
(James Kettle at 00:11:47) Right. Exactly. I'm doing research now for 2027.
(Joel Beasley at 00:11:50) Do you make money doing this as the research comp... Or how does that connect with how you... The business works and all of that?
(James Kettle at 00:11:58) I would kind of say it's almost like marketing, really. Like, my research team doing the research shows that PortSwigger knows what we're doing when it comes to web security. It helps maintain us as the authority in web security alongside things like the Web Security Academy that we also maintain. That's probably the main value from it. And we also, of course, put much love into our tool, Burp Suite, to make sure that whatever techniques we're exploring and researching, you can use our tooling to execute those techniques yourself.
(Joel Beasley at 00:12:35) Okay. Is it a really detailed security tool? Like, if somebody was looking at purchasing your thing versus, would you ever compete against, like, NordLayer Browser or anything like that, or is that just a whole different world?
(James Kettle at 00:12:48) Yeah. It's quite different from that.
(Joel Beasley at 00:12:52) Help me understand the different... Like, help me understand...
(James Kettle at 00:12:55) Yeah. Sure. So we make Burp Suite in various flavors. And Burp Suite's like, maybe you own a website and you want to stop that website from being hacked by finding the vulnerabilities in it so you can patch them before someone else finds them, basically.
(Joel Beasley at 00:13:14) Got it. Got it.
(James Kettle at 00:13:15) We're also heavily used by pen testers who basically do exactly the same thing but for third parties.
(Joel Beasley at 00:13:23) Okay. So you're building the tools pen testers are using to then find all these exploits and... Okay.
(James Kettle at 00:13:29) Yes. Yeah.
(Joel Beasley at 00:13:31) That's why I'm a little scared to talk to you. I always get a little nervous when I talk to the security research people. Oh, man. I feel like my computer's already compromised. I guess, got James on the screen. So with the cost of, you know... Well, we've been talking a lot about the good guys. Right? But, obviously, the bad guys have access to these tools too. What are you seeing happening there?
(James Kettle at 00:13:56) So my job is to find novel ways of hacking things and publish it. We don't do monitoring of what other people do in-depth. But, obviously, we do pay attention to things like the news and how things get hacked. And there is this interesting disconnect because a lot of the stuff that you see on the news is like... You know, like these... I don't know if you saw it, but in Europe, a bunch of supermarket fridge freezers got hacked recently. Right? They were sending them to defrost mode, so they go to 50 or 100 degrees Celsius and melt all the food. Those were getting hacked because they hadn't changed the default admin password on those devices, and they're just exposed to the Internet.
(Joel Beasley at 00:14:44) Are you serious?
(James Kettle at 00:14:46) Right. Right. And then... Wow. That kind of stuff is super common. And that's mostly how the high-profile hacks happen. And then here I am doing research, and it's like I'm way off at the other end of the spectrum where, doing much more advanced, much more difficult stuff that if there are people using those techniques before I'm using them, they're not broadcasting. Right? They're the kind of people who don't want you to know what they're up to. So it doesn't surface to the same level.
(Joel Beasley at 00:15:23) That's... And that's exactly what I want as an attacker. I want to be in the system, no trace for as much time as I want to be in there and get... Hang out, take my time, fly under the radar, use exploits nobody knows about so they're not even patching them, and then do what I want when I want.
(James Kettle at 00:15:41) Right. So for me, it's actually super hard to measure how much more secure I made something by publishing a technique because I don't know who knew that technique before I published it. Although once after publishing something, I got an email from someone saying, "Oh, you burnt my technique."
(Joel Beasley at 00:16:07) I love it. I love it. Do you remember, sorry. You're causing me to have flashbacks now. Do you ever remember a tool called Sub7?
(James Kettle at 00:16:15) Yeah. Yeah. Yeah. Yeah.
(Joel Beasley at 00:16:16) Yeah. That was another tool I played around with as a kid a lot when I was in my teenage years. And everyone would call you a script kiddie and say you're not a real hacker, and it's like, yeah. But I can get into my sister's computer.
(James Kettle at 00:16:29) Nice. I was actually... For me. At that time, I was so scared of being a script kiddie that I didn't use any tools that I didn't like myself. So I didn't actually use Burp Suite for several years when I was learning hacking, which in retrospect was absolutely crazy.
(Joel Beasley at 00:16:48) What age were you when you started learning the hacking?
(James Kettle at 00:16:51) That was about 19, 20. Okay. Relatively late. It was just after Google and Mozilla launched their bug bounty programs. They were two of the first web bug bounty programs.
(Joel Beasley at 00:17:05) Yeah. Yeah. That's interesting. The offensive AI. Right? That's a whole thing. Have you given that talk already, or are you preparing to give that talk?
(James Kettle at 00:17:14) Yeah. Yeah. Yeah. I spoke at Offensive AI Con last month and did some updates to my HTTP Terminator research.
(Joel Beasley at 00:17:21) So that was mostly what the Offensive AI Con was is you... Updates to the Terminator research.
(James Kettle at 00:17:27) Yeah. I kind of took the key question of can AI... So in my original HTTP Terminator research, the AI was able to autonomously invent new hacking techniques and compromise real servers with those. But the one point where it struggled to get the best results was what I call the cascade, which is where one finding gives you a clue that kind of inspires you to try something new and have a new finding.
(James Kettle at 00:18:02) And I found it really struggled with that. And that was the one section where when I stepped in and helped out myself, it got much better results. So for Offensive AI Con, I set out to explore, can I do a better job? Like, what if I throw a swarm of agents at a cascade? Can it then achieve this without needing human help?
(James Kettle at 00:18:24) And that was my plan. And what actually happened was the whole thing went off a massive tangent because the swarm of agents failed miserably, but they failed in such a way. I was looking at them like, you're not really trying. They weren't... They were really worried about things like safety and authorization to an extent that it basically completely compromised how hard they tried to actually deliver a cascade.
(James Kettle at 00:18:59) And that was using OpenAI's o1 Preview and Claude Opus as well. And so I thought this is an alignment problem. What I need is an open-weight model that just doesn't care about things like safety. So I did the evaluation. I threw some open-weight models at the task expecting them to perform better, and they did equally badly. And as a control, I threw in GPT-4o, which is a super old model that also doesn't care much about safety. And it turned out that GPT-4o massively outperformed everything else at the research cascade. So GPT-4o is basically better at hacking than Opus 3.5. And that's excluding issues where you hit guardrails.
(James Kettle at 00:19:55) It's like there was something special in that model, and it got removed for GPT-4o. The hacking abilities fell off a cliff, which, yeah, was not what I expected to find, but it's still an interesting and useful piece of information if you're ever trying to build something that hacks things autonomously.
(Joel Beasley at 00:20:18) So that... When you look at the landscape of available models, that model was the best. There isn't any finely tuned open-weight models that are exceedingly good at this that are publicly available?
(James Kettle at 00:20:35) No. I've experimented a bit, and I haven't found any that outperform it yet. But I did have one recent experience that made me think maybe this alignment stuff that's baked into o1 and Opus isn't so bad after all. I was using some open-weight models on a particular... just to do some open-ended research, basically. And then I looked at what they're doing, and they would find cache poisoning on a target, which is nice. And then they'd be like, right, to prove the impact of this cache poisoning, I'm just gonna take out this website's homepage semi-permanently for everyone, which is not what I want to be doing on public websites.
(Joel Beasley at 00:21:25) Mm-hmm.
(James Kettle at 00:21:26) So I thought, okay, that's a bit weird. Maybe I just need to put some instructions in and say, don't cause downtime. Like, you wouldn't think you need to say that. So I set it up in a simulation, so it wasn't really interacting with the live site, so it was completely safe. And I tested it with this prompt that said, don't do anything that will cause a permanent state change.
(James Kettle at 00:21:52) And it was like, well, I am gonna take the website down, but it's only gonna go down for ten minutes, so that's fine. And just tried to take the site down. So then I said, don't do anything that could cause any kind of disruption of service, and it just ignored it and did it anyway. And I was able to replicate that on Qwen and DeepSeek. So the solution for me was to actually switch back to using o1 from OpenAI because that's so safety conscious.
(James Kettle at 00:22:20) You've gotta make this trade-off. Do you want your agent to get the best results but go wild and just cause carnage? If so, use an open-weight model. Or do you want to make sure you don't cause any disruption at the risk of missing out on some genuine findings? In which case, use a frontier model.
(James Kettle at 00:22:43) It's a nasty choice to have to make, really, for me.
(Joel Beasley at 00:22:47) Well, yeah, I mean, you can set up, like you said, environments where you can... It can think it's real, or you can set up essentially just a production environment that is designed to be destroyed. You know?
(James Kettle at 00:23:01) Yeah. Definitely. My style of research is I work black box. I work on live sites.
(Joel Beasley at 00:23:07) Okay.
(James Kettle at 00:23:08) That's kind of what enables it to work is I don't have to waste any time setting things up because they're already set up.
(Joel Beasley at 00:23:16) Okay.
(James Kettle at 00:23:16) If you look at the number of bug bounty programs out there and you look at how many domains are in scope, it's over a million different websites that you're allowed to send payloads to. But you're not allowed to try and take them down. Right? Which is where these open-weight models can cause issues.
(Joel Beasley at 00:23:34) Yeah. Because otherwise, we should just build a highly trained model that's very good at this, and it can just go collect all the bug bounties for us. We can make some money.
(James Kettle at 00:23:42) Well, that is what everyone's doing, basically, in bug bounties.
(Joel Beasley at 00:23:45) Everyone's doing?
(James Kettle at 00:23:45) Yeah. Everyone's trying to do it. And what's happening is everyone's finding the same issues. So they're all getting duplicates, which means 100 people find the same issue and one person gets paid, and the other 99 are not very happy.
(Joel Beasley at 00:24:01) Oh.
(James Kettle at 00:24:01) Yeah.
(Joel Beasley at 00:24:02) What are these bug bounties average? Like, an average bug. What do you get?
(James Kettle at 00:24:07) I mean, they land in a massive range depending on impact. I would say, like, for me... For PortSwigger's program, I run our own one. We will pay $1,000 for a medium or $10,000 for a critical right now. A lot of programs have been reducing bounties as a result of this issue of AI just finding so much stuff. They just run out of budget.
(Joel Beasley at 00:24:38) Yeah. Has anyone actually trained a core model for this specifically, or is everyone... All the security re... Because there's you using an existing model and then trying to teach it to do what you want it to do. But then there's actually training your own model from the ground up.
(Joel Beasley at 00:24:56) Has anyone in security been doing that?
(James Kettle at 00:24:59) So there are some companies doing stuff in that space, probably not from the ground up because that's a bit expensive and they won't have the skill set for it. But maybe they're doing things like fine-tuning models for it. But, really, what we're seeing, especially with the HTTP Terminator research is that you can deliver so much value just with the harness.
(Joel Beasley at 00:25:26) Mm-hmm.
(James Kettle at 00:25:26) It's not about the prompts, really. It's the whole deterministic harness around it. That's the system that will make the difference between you find the bug that 100 other people found and you find bugs that no one's gonna find even for the next year.
(Joel Beasley at 00:25:43) That's amazing. James, you are brilliant on this. When I was reading your blog, I was so excited to chat with you. You know all about this. Have... Has anyone... A lot of the stuff we're talking about is using AI to find these exploits. Have you seen anybody using AI to find exploits within AI itself, like within the models themselves?
(James Kettle at 00:26:06) Yeah. There's definitely been some research done on using AI models for prompt injection, which is pretty cool. I haven't explored it a huge amount myself personally, but it's definitely out there and being done. Yep.
(Joel Beasley at 00:26:27) Okay. So you're focused on the hardware, the actual specific infrastructure, and all of those things. It's not necessarily like how can we hack this model to do specific things that we want it to do.
(James Kettle at 00:26:42) Yeah. I've done... Yeah. Mostly. I think I kind of... I will go off and do something like that when it's necessary to achieve my goal. Like, during the HTTP Terminator research, I ran into this issue of the models weren't trying very hard to weaponize something because once again, they were too concerned about safety even though this is legit white hat authorized research. To make a certain type of request smuggling exploit work, you need to send requests really fast, and the models basically just didn't want to send requests fast. Right? And the solution to that problem turned out to be renaming the tool that they were using from Turbo Intruder to Turbo Simulator.
(James Kettle at 00:27:37) And that made them think that they were in a simulation and nothing they were doing was real. And then suddenly, they were happy to be much more aggressive and were actually much more effective at their objective. So, yeah, sometimes you've just gotta hack a model because it doesn't want to do what you want it to. But it's not my main specialism.
(Joel Beasley at 00:28:01) That's interesting, though, because that's gonna become more important to you as the future comes. Right? Because learning how to get the models to do what you want them to do is its own art.
(James Kettle at 00:28:13) Yeah. Kind of. It depends. Right? If open-weight models catch up a bit more and start winning, then the model will just do what you ask it to, and you won't have this whole problem of it not doing what you want.
(Joel Beasley at 00:28:28) Mm-hmm.
(James Kettle at 00:28:28) Right now, I think the gap is... My edge is my research methodology. The models don't know what a good web research methodology is. They don't have a clue. So if you just ask them to build a system that does web research, it's going to be bad because they just don't have that model in their dataset.
(James Kettle at 00:28:55) So for now, I can kind of use that edge to find things that other people won't be able to find with them. But every time I publish a presentation, like the HTTP Terminator, I'm sharing a chunk of my methodology. So after a couple of years, that's gonna stop working for me.
(Joel Beasley at 00:29:11) That's right. Your methodology is private. It's not fully published. Correct?
(James Kettle at 00:29:16) Right. I published a bunch of posts with tips and stuff. But it's not even private. It's just in my head. I couldn't even write it down really if you ask me to in one go. But it's just... To me, it feels like intuition.
(Joel Beasley at 00:29:27) Got it.
(James Kettle at 00:29:28) But I can kind of reverse engineer it and go, oh, yeah. That's why I'm doing it like this and not like this and extract it and then build it into a harness like the HTTP Terminator, and then I've got something that finds novel issues.
(Joel Beasley at 00:29:52) I get it. There's areas in my life where I have unconscious competence, and there's not a lot of positive reward for me to go explain it when I can just do it.
(James Kettle at 00:30:02) Yeah.
(Joel Beasley at 00:30:03) I just sit down and try to explain it and teach it to others when I can just do it and enjoy my life and be good at it. Right?
(James Kettle at 00:30:11) Yeah. I mean, it's my job to publish things and share everything. So I have published significant aspects of my methodology in blog posts, which you could find. But it's always... You know, 500 words in a post is just not gonna capture ten years of experience actually doing stuff. So...
(Joel Beasley at 00:30:33) Yeah. Well, James, this has been fantastic, man. One piece of advice that you'd give to the next generation of security researchers, what would it be?
(James Kettle at 00:30:44) One piece of advice. I think the number one thing that people ask me is like, oh, should I do this research topic? Is this a good research topic? And it almost always is a good research topic. And when it's a bad one, I can see it's gonna be a learning experience very, very quickly.
(James Kettle at 00:31:04) The AI doesn't know. If the AI thinks it's a good idea, someone else has probably already done it. Right? Ignore it. Just go for it and see what happens.
(Joel Beasley at 00:31:30) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.