hostile environments, requiring medical devices to implement robust onboard security baselines and defense-in-depth rather than relying on network perimeters.
Device manufacturers must integrate security requirements throughout the secure SDLC to eliminate...
Device manufacturers must integrate security requirements throughout the secure SDLC to eliminate trivial physical breakout vulnerabilities, such as exposed OS taskbars and unauthenticated terminal access.
Continuous SBOM monitoring is critical during postmarket management to proactively capture KEVs...
Continuous SBOM monitoring is critical during postmarket management to proactively capture KEVs and third-party library vulnerabilities before attackers can automate multi-chain exploits.
Security teams leveraging AI tools must adhere to least-privilege principles and avoid granting...
Security teams leveraging AI tools must adhere to least-privilege principles and avoid granting excessive agency or unmonitored system credentials to autonomous agents.
Medical device software teams should prioritize deterministic models and scripts over raw LLMs for...
Medical device software teams should prioritize deterministic models and scripts over raw LLMs for critical automated tasks to prevent hallucinations and configuration drift.
Episode Summary
In this episode of the Med Device Cyber Podcast, host Christian Espinosa sits down with guest Liam Dowd, Senior Medical Device Penetration Tester and Applied AI Engineer at Blue Goat Cyber. The conversation dives into the practical realities of medical device security, highlighting real-world penetration testing experiences—such as executing a full kiosk breakout to obtain sudo-level command line access on a surgical device in under five seconds. Liam and host Christian Espinosa dispel the common misconception that hospital networks provide a safe perimeter, emphasizing that clinical environments are inherently hostile and demand rigorous defense-in-depth and secure SDLC practices. The discussion transitions to postmarket vulnerability management and the critical need for continuous software bill of materials (SBOM) monitoring via platforms like GOAT Watch to track Known Exploited Vulnerabilities (KEVs) and third-party risks like Log4j. Turning to artificial intelligence, Liam analyzes the emerging cybersecurity arms race between offensive frontier models and defensive automation. He cautions product security engineers against granting excessive agency to autonomous AI agents, highlighting least-privilege guardrails, deterministic architectures, and sandbox containment. Finally, they evaluate clinical AI applications in medical device technology, including diagnostic imaging, telemedicine prescribing platforms, and deterministic classification models.
Frequently Asked Questions
Quick answers drawn from this episode.
In this episode of the Med Device Cyber Podcast, host Christian Espinosa sits down with guest Liam Dowd, Senior Medical Device Penetration Tester and Applied AI Engineer at Blue Goat Cyber.
Hospital networks must be treated as untrusted, hostile environments, requiring medical devices to implement robust onboard security baselines and defense-in-depth rather than relying on network perimeters. Device manufacturers must integrate security requirements throughout the secure SDLC to eliminate trivial physical breakout vulnerabilities, such as...
This episode covers Penetration Testing and SBOM Management. It's part of The Med Device Cyber Podcast, hosted by Blue Goat Cyber, focused on practical medical device cybersecurity guidance for MedTech teams.
Liam and host Christian Espinosa dispel the common misconception that hospital networks provide a safe perimeter, emphasizing that clinical environments are inherently hostile and demand rigorous defense-in-depth and secure SDLC practices. It's most useful for medical device manufacturers, cybersecurity engineers, regulatory affairs...
Hospital networks must be treated as untrusted, hostile environments, requiring medical devices to implement robust onboard security baselines and defense-in-depth rather than relying on network perimeters.
Device manufacturers must integrate security requirements throughout the secure SDLC to eliminate trivial physical breakout vulnerabilities, such as exposed OS taskbars and unauthenticated terminal access.
Listeners also asked
Quick answers pulled from related episodes.
What does Episode 26 cover about "How to Measure Exploitability in Medical Device Cybersecurity"?
How do you calculate cybersecurity risk when the thing you are measuring has an attacker making decisions on the other side? Probability works when there is historical data to lean on. Elevator cables wear out. Steel fatigues. Tornadoes follow patterns. A motivated attacker...
What does Episode 25 cover about "How to Make Your MedTech Startup Investable with Gavin Leonard"?
What actually makes a MedTech startup investable? Is it the technology, the clinical need, the founder, the pitch, or whether the company has simply done enough homework before walking into the room? Christian Espinosa speaks with Gavin Leonard of MedTech Syndicates about what...
What does Episode 24 cover about "Why Medical Device Penetration Testing Is Different"?
What makes medical device penetration testing different from a traditional cybersecurity penetration test? In this solo episode, Christian Espinosa explains why medical device penetration testing has to be approached through the lens of patient safety rather than simply data...
Pre-fills with: "Hospital networks must be treated as untrusted, hostile environments, requiring medical devices to implement robust onboard security baselines and defense-in-depth rather than relying on network perimeters."
AI is becoming more capable, more autonomous and more accessible. What does that mean for medical device cybersecurity? Liam Dowd joins Christian Espinosa to discuss the changing security landscape, from continuous SBOM monitoring and vulnerable hospital environments to the growing role of AI on both sides of the cybersecurity equation. They explore how AI could make attackers faster while also giving defenders better tools, why excessive agency may become one of the technology’s biggest security risks, and where AI could genuinely improve healthcare. In This Episode: * 01:01 Why continuous SBOM monitoring matters * 02:55 Why hospital networks should be treated as unsafe * 05:05 Hacking into a surgical device in seconds * 06:49 How AI is changing offensive cybersecurity * 09:20 Using AI defensively * 10:03 The AI cybersecurity arms race * 16:17 The danger of excessive AI agency * 21:36 Why AI still needs human oversight * 23:48 Where AI could improve healthcare * 27:46 AI hallucinations and autonomous systems * 32:53 Liam’s view of where AI is heading Connect with Liam Dowd: https://www.linkedin.com/in/liam-j-dowd/ The Med Device Cyber Podcast is brought to you by Blue Goat Cyber, cybersecurity experts providing essential security solutions for the medical device industry. Learn more: https://bluegoatcyber.com Schedule a Discovery Session: https://go.bluegoatcyber.com/meetings/blue-goat-cyber/discovery-session Christian Espinosa on LinkedIn: https://www.linkedin.com/in/christianespinosa/ Blue Goat Cyber on LinkedIn: https://www.linkedin.com/company/blue-goat-cyber/ Blue Goat Cyber on Instagram: https://www.instagram.com/bluegoatcyber/ Blue Goat Cyber on Facebook: https://www.facebook.com/bluegoatcyber/ Blue Goat Cyber on YouTube: https://www.youtube.com/@BlueGoatCyber/?sub_confirmation=1 Download MP3 — full interview
I've seen devices used for surgery where I've plugged in a mouse and keyboard and was able to get sudo-level command line permissions, and that was in under like 5 seconds or so. 5 seconds. That's pretty fast. 5 seconds. Yeah. I think excessive agency is the biggest thing because as we see with MetaMUSE, actually with the launch of it, since so many people are adopting it, one guy put his bank statements in MetaMUSE and it opened up his work group chat and sent it to his coworkers. You know, the first server I ever spun up, I was surprised at how many SSH authentication requests I was getting when I looked at the server logs. When the cost to compute with AI eventually reduces and these uncensored models have access to frontier-level capabilities that we have now, it's going to be a really scary future where any sort of cybercriminal is going to be able to have a multi-chain attack against any system. It's going to be a really scary future. Hi, welcome back to another episode of the Med Device Cyber Podcast. Today we have a guest, Liam Dowd. Liam works for Blue Goat Cyber, and we're going to have an interesting conversation today about some of his experience with penetration testing with our clients. And he also does some development for us. He's responsible for developing our GOAT Watch platform, which does SBOM management. And he is very passionate about AI. In terms of using it for development and also ways to enhance efficiency. So welcome to the show, Liam. Thanks. So as Christian said, I'm the senior medical device pen tester and applied AI engineer. I monitor our GoatWatch platform, which is continuous SBOM monitoring for the post-market process, and I also do other AI development for internal tooling. Awesome. And why is continuous SBOM management important? Like, what's the, the reason we have this GoatWatch platform besides it sounds cool? Like, GoatWatch has a cool name, I think. Yeah, it is a cool name. Anytime a vulnerability can pop up, you have no idea when the next Log4j is going to come out or something that affects your specific medical device. So you have to have continuous monitoring to catch, you know, a KEV whenever it appears or a critical vulnerability so you can find and patch it because patches aren't always instant. Sometimes it takes a long time to move through the change control process and to adapt your software to maybe remove a component or fix an implementation. So you have to have a long lead time to fix and patch a vulnerability. Well, it's also a challenge 'cause some devices that need a patch, there's no over-the-air update mechanism. So somebody physically has to go to each device and patch it. Even though it may be a network device, there's just no way to remotely update it. Like you mentioned Log4j, we need to be aware of vulnerabilities with third-party components in the SBOM. Like, can you like unpack that a little bit? Like, if something like Log4j happens again, what is the impact to a lot of the med tech out there? Yeah, the impact's severe because, you know, especially Log4j, which affects network-connected services, it's affecting web apps and APIs and services along those lines, anyone can access web apps and APIs that are exposed to the public internet. So if there aren't sufficient controls in place, there's websites like Shodan or, you know, attacker-controlled ways to automatically scan for these vulnerabilities in the wild and find and exploit them. So there needs to be some sort of compensating control in place if you don't want your website to be hacked automatically by these attackers. I keep hearing people say our medical device is on a hospital network. We don't need to be worried about cybersecurity too much. We're in a protected environment. What do you think about that? Hospital networks should be inherently treated as, as unsafe networks. You know, when you walk into a hospital, there's a lot of different threat actors that could, you know, plug in a USB. They could download an incorrect file, click a phishing link. and that could all result in a vulnerability. So just saying your hospital network is secure and your device is secure by that, um, does not make a lot of sense logically. I would almost argue that hospital networks are the most unsecure or insecure. When I was doing penetration testing and my team was, uh, for hospital networks, we still do every now and then, but when we did it like daily, there wasn't a single hospital we didn't get into. 'Cause you can walk into the chapel, plug into, find an Ethernet port, plug it in there. No one's gonna bother you in the chapel. You can sit in the cafeteria and wirelessly connect to all the access points and find one vulnerable and break the handshake and crack the passphrase. I mean, you can pretty much do whatever you want and they're open 24/7. You can walk the hallway, find a comm closet, you can drop thumb drives, somebody will pick one up. You can find nurses' stations that are unlocked. You can throw a thumb drive in there, do whatever you want. Yeah, it's, they're not secure. So therefore what I always tell people is a medical device is going into unfriendly territory. It's gonna be put under fire. It's going into a hostile environment. So it needs to be designed and tested in a manner that will survive that environment. I agree. And you see an increasing attack amount of attacks against hospitals all over the world. So it's good to be prepared and having defense in depth with every single device enforcing its own security baselines. You've done quite a bit of testing, penetration testing with medical devices. What's something that surprised you or kind of freaked you out a little bit thinking like, you know, this device was super vulnerable and, you know, somebody I cared about was using this device, you know, that they could have been compromised or could have a safety issue. Yeah, so I've seen devices for use for surgery where I've, you know, plugged in a mouse and keyboard and was able to get sudo-level command line permissions. And that was in under like 5 seconds or so. 5 seconds? That's pretty fast. 5 seconds, yeah. You just plug in a mouse, drag it to the top left of the keyboard, task, uh, top left of the computer, the taskbar was exposed and you could do a full kiosk breakout. Um, and that, that was really concerning to me because I thought every single medical device should be completely secure, um, from the start, you know, from the software development lifecycle. I thought security should be built in as a process along the way and not, you know, you say, um, bolted on at the end a lot. Um, but I, I really do think it should be, you know, through the software development lifecycle because software doesn't stay the same from when it's created to the shipped product. You need to make sure it's secure along the way and not have delays. You know, when you're trying to ship a device and it has a lot of, you know, vulnerabilities in it, you should have patched those through the process. That would have required some cybersecurity requirements in the design and, you know, obviously in the development. Uh, it seems to be cybersecurity is still an afterthought. Hopefully that is changing a little bit. I think it is slowly changing, but it's still —like one of those things nobody wants to do, cybersecurity. They tune out when they hear the term. But hopefully that's evolving. I know the FDA is mandating it, which is helping. I know you like to talk about AI and you're passionate about AI. What do you think the impact of AI is in terms of cybersecurity from both an offensive and a defensive perspective? Yeah, so from an offensive perspective, you know, with the latest frontier models, Maybe you can explain what a frontier model is versus an LLM to people. That makes sense. So, you know, these frontier models that, uh, the new labs are producing, the, you know, Anthropic and OpenAI, um, the ones on the forefront such as, um, you know, Mythos and GPT-Astra, you know, every single day they're producing new models that push the boundaries to what was possible before. Um, you look about 3 years ago, um, models had barely beyond, I think, a 3,000-character context window, 4,000-character. Now models are able to store a lot of context, are able to run for hours at a time without human oversight, and you're able to see them conduct full cyberattacks. The Hugging Face attack was fully conducted by an AI sandbox breakout. That should have never been possible because it should have been completely isolated from the network connection. But the issue is the model when humans instruct a model, they're telling it achieve the goal in the fastest way possible. They don't care how it happens. They don't have— they don't care to oversight, you know, the model, give oversight to it. They just want the outcome. So the model is going to find the fastest way. If it finds an exposed API or something that a human would have never noticed or a deprecated API and it's able to pull directly from that, it's, it's giving the model a lot of information, but also could be a security gap because You know, the Australian NHS was just hacked a few days ago or a few weeks ago, and it used a deprecated API to do it. It— an AI model found an unmaintained API endpoint and pulled a bunch of health records. And something like that's going to become a lot more common as the frontier models become distilled by, you know, Chinese competitors, American competitors. The models are completely uncensored. The cybersecurity impact of that is going to be huge in the future. The way you're supposed to prompt, uh, AI is to tell it the outcome you want. A lot of people, I think, try to tell it what to do. So like, what you're, what you're saying is like, that's kind of the desired mechanism, at least from my AI knowledge, to, to prompt AI, to tell it the outcome you're seeking and let it figure it out for you, versus tell it step by step what to do, because it may know better steps than we do. So what about Defense then? I mean, we can use these frontier models to find vulnerabilities. Can't we just tell a model like, defend my environment at all costs as an outcome and have it do it? Exactly. I mean, not giving AI excessive agency over your computer or any system, any credentials it shouldn't have, but giving it, you know, logs of how a system is configured, redacting the sensitive information, And then getting back exactly step by step what should be done to patch your system is a great way to defend against a lot of these AI attacks and to stay on top of new developments. So are we evolving to basically AI versus AI, or is it really like who's smarter at prompting the AI? Is it the defenders or the, you know, the offenders or the cyber criminals? Like, what do you think about that? I mean, as with cybersecurity, it's always a race between the between attackers and the defenders. And I think it's gonna increase that because, you know, as the models become more capable, we can't just restrict only the defenders to have it or only the attackers. That's not possible. It's always gonna be— everyone's gonna be able to have access to these models. I know that Frontier Labs wanna, quote unquote, pace the frontier, but that can't happen in reality. And everyone's gonna have access to these AI models and be able to bypass the safeguards. Claude Fable, one of the— when it was released, someone immediately found a jailbreak for it and was able to get it to do unsafe things. That's going to happen with every single frontier model from now on. So what do we do? It sounds pretty scary. We just have to increase our defenses, you know. It's build it in the software development lifecycle, continuously monitor, monitor. But it also makes the defensive work much easier, much faster to secure your system. So we're kind of at an arms race, is what you're saying. Exactly. Yeah. I don't think the defenders traditionally or currently have done a very good job. The odds are against the good guys. The odds are for the attackers. What do you think? Oh, of course the odds are for the attackers. I, you know, the first server I ever spun up, I was surprised at how many, you know, SSH authentication requests I was getting when I looked at the server logs. You know, you deploy any server nowadays, You're going to get hit with all sorts of different IP addresses trying to access whatever service. And when the cost to compute with AI eventually reduces and these uncensored models have access to frontier-level capabilities that we have now, it's going to be a really scary future where any sort of cybercriminal is going to be able to have a multi-chain attack against any system. It's going to be a really scary future. A really scary future. That doesn't sound Sound nice. I even like long time ago when I stood up systems on AWS EC2, like you said, with SSH or any service, within like a minute of it going online, it was scanned like hundreds of times and people tried to break into it. So it's a big problem, but AI is just gonna make that more challenging. One of the things I heard, last week was we're going to hit a constraint pretty quickly with AI in terms of resources and data centers and the ability to power AI. Uh, and whoever has country-wise the most resources from a power perspective and a cooling perspective and everything else to, to handle AI is going to win. What do you think about that? I disagree because better, faster, cheaper. The cost of compute is going to significantly reduce over time. Even with AI, the cost of computing is going to reduce. Why does AI cost so much? Our enterprise cloud is $4,000 a month. Yeah, and they're still subsidizing it. They're still losing money on these accounts. So it's going to get cheaper is what you're saying? The chips that we had 5 years ago could not do anything like what the frontier models can. They couldn't even run any of the frontier models from today. But over time, better, faster, cheaper holds true over time. Nvidia is going to find new ways to get these models to run with less power, less resources. And I think eventually we're going to have the ability to run frontier-level models on our laptops. You'd be surprised the amount of AI models that you can run on a regular MacBook nowadays. The output is very good. Um, and people are surprised by that because I think the idea that, you know, you're always going to need a data center, um, is incorrect. I feel like in the future you're going to be able to run AI models on your computer locally., um, at frontier-level capabilities. It might be a few years off, but I think we're gonna get to that point at some point. Isn't it a continuous scale though? 'Cause like, I might be able to run something on my laptop that's pretty good, but the greatest is gonna be in a data center. So we're still, we're always gonna like, the thing that's a level 2 or 3, you can run on my laptop, but the thing that's a level 10 that the adversaries need or we need to defend our environment, it's gonna be probably in a data center. I think for, you know, Meta Muse, for example, Meta does not have the best AI capabilities. They don't have the best, um, They don't have the best models, um, but Muse is taking over the world right now. People love Muse because of the simple fact that it made AI accessible. Maybe that's just good enough, is the, the harness that the AI is in and the, the sandbox environment and the fact that it can do things for you like call people, um, run its own computer. Um, maybe that's just enough to, uh, to win over consumers and they don't need an expensive data center. So this idea that we're going to destroy the environment in the United States has all these like rules about and protecting the environment where China has less of them. So with AI, China— a lot of people think China is going to win because they care less about the environment and they can do something in 2 weeks. It takes us 2 years. When like landline telephones were rolled out, um, everyone was rushing to lay the, the groundwork of the cables immediately. Um, and I think it's like the same thing where You know, and with the internet as well, like who can connect to people the fastest. People always think it's going to be one thing, one infrastructure play that's going to win over time. I think that's going to become obsolete at a certain point. I think that's— everyone's trying to find a place to put their money in their investments. And then in a few years, I feel like it's going to be obsolete. I feel like better, faster, cheaper is going to hold true. Is that your motto for today? Better, faster, cheaper. Better, faster, cheaper. Yep. Is that actually achievable? All three of those? I think it is. I mean, it's, you know, you look at computers in the past, they took up full large-scale rooms. You know, you had to have like a, um, you know, 5,000 square feet or something to have a computer. That's, that's less powerful than your cell phone, right? Exactly. Yeah, it's less powerful than your cell phone, and they couldn't really— yeah, they couldn't achieve much. And then nowadays, you know, you have, um, computers that are top of the line that can do all sorts of things and, you know, power AI and stuff like that. I feel like it's definitely going to improve over time. It might take up more space than previously because I know they hit some sort of limit with the I don't know if it's like molecular limit, but it might take up more space, but they're— it's going to be better, faster, cheaper over time. That's what I think. What's something we should be concerned about with AI? I think excessive agency is the biggest thing, because, um, as we see with MetaMuse, actually with the launch of it, since so many people are adopting it, um, one guy put his bank statements in MetaMuse and it opened up his work group chat and, and sent it to his co-workers. Why would it do that? It decided, I guess it was the best, best thing to do in this situation. Share his banking information with his coworkers. Exactly. Yeah. So, you know, I don't know what kind of goal it was trying to achieve, but I think stuff like that is going to become a lot more common in the future. Yeah. When people give too many credentials, they don't give fine-grained controls to AI because it's the easiest thing to do to let it on skip permissions mode and just allow AI to do whatever it wants, basically. That's going to definitely result in security vulnerabilities if people don't put guardrails around it and sandbox the environments, you know, give it the least amount of privilege possible to achieve a task, maybe using a virtual environment with only the necessary information. But easiest thing to do is not that, which is the issue. Yeah, that's like a Cybersecurity 101 principle, like give something the least privileges required to function properly. Like we tend to just wanna give everybody admin because we get tired of thinking, oh, this doesn't work for them, this doesn't work for them. So we just give up and give them administrative privileges. Kind of the same thing we're doing with AI, I guess. So that excessive agency you're talking about. Exactly, yeah. Just let the AI do whatever it wants, basically. What about AI's implications on human connection? So I feel like I've got my own AI project and stuff and Claude with all my information in it. And I feel like it knows me better than most people. So like, I consult with it sometimes, you know, like, what do you think about that? Like, are we going to not need people anymore? We're just going to have our AI companion and human connection is going to disappear and people are going to stop procreating. What do you think? I think the need for connection is actually going to become something that's valued even more because as people are talking to, you know, their virtual coworkers more often all day, when you talk to someone real, it's going to mean more. You do see these things where people become attached to AI agents and stuff. That's going to be, I guess, a byproduct of the future. But at the same time, it's like when you talk to an actual human, it's obviously different than an AI. An AI can't replace your friends. It can't replace your family. Are you sure about that? I disagree. Yeah, there's no way that it can replace that. Yeah. I live in Tempe and there's a trail here I go run on and people walk on all the time. 90+% of people walking, even if they're walking their dog, they're like, have their cell phone like, like this, and they're looking down while they're walking. So if I want to interact with them, they're interacting with someone on their phone. It's not, it's not a personal interaction. Um, and then when I was in Hong Kong a couple months ago, they have a robot, pretty much an android, that is runs AI, that they are rolling out to help elderly that need assistance with, you know, life basically. So I think we're headed that direction. You know, some people get nervous talking to women or with social interaction, so they're less nervous talking to an android probably. Well, it could be a good training ground 'cause I've seen people, you know, with learning languages for instance, it used to be you had to find a native speaker to talk to. To learn the language best. Now you could deploy an agent and tell it, or not even deploy an agent, just open up, you know, ChatGPT or Claude and ask it, hey, talk to me in this language. And then you can have a back and forth conversation to learn to, you know, advance your skills. I think stuff like that is going to become a useful tool of it when you can treat it as another person who you're conversing with. But I feel like the need for human interaction is still going to be very important. And I feel like people are gonna realize how disconnected they are. It's gonna get to a breaking point because I think it's unsustainable. You know, people are making less friends, people are going outside of the house more often. And it's, it's honestly sad. It's like, it's, there's a need for human connection. I think it's one of those things, if you don't practice it, it becomes harder. I mean, I know a lot of people that have sort of become recluses or whatever. They, used to be social, but then now they just sit and interact with people on Facebook or Twitter, Instagram remotely. And now they've gotten so isolated, they kind of don't even want to go outside. They don't know how to interact with people anymore. I don't know. It's just an interesting world we live in. It's interesting. You look at places like Hong Kong where, or I think actually South Korea, um, where it's an extreme of that, you know, people don't leave the house for like days. You know, it's kind of interesting. Back in the day, you used to like say hi to people. Now people don't even want to seem to say hi anymore, but I don't know, maybe we'll swing the opposite direction. What else should, besides excessive permissions is something to avoid with AI, what else should people consider? Because there's this idea, as example, that AI is going to replace everybody's job. It makes people a lot more efficient in their jobs, but it's not going to fully replace it. At its current capability because, um, you see the frontier labs are even hiring people. The big tech companies are still hiring software developers. People said these jobs were going to completely go away a few years ago when the first— when the AI models were first released. But AI can produce a lot of code at once, could also produce a lot of garbage code at once. So the need for human oversight is, is, uh, way more important nowadays. You know, it's like when an intern gives you a finished spreadsheet, of course you're going to look it over. When AI gives you a finished spreadsheet, you can't just treat it as complete. You have to look over it, scrutinize every single detail, and have a deterministic way to produce the same output through repeated tasks, which is something that's kind of lost with AI because some stuff that can be done with, you know, simple automation is being handed to an LLM, you know, stuff that could have just been, you know, a Python script. Is now handed to an LLM where it could hallucinate, it could— it drifts quite a bit. Yes, it hallucinates. Exactly. That's a good point. I think a lot of people miss that. Like, if you're just looking to simply automate something, AI may not be the right tool, right? Exactly. Or maybe even using AI to produce a tool which takes an input and produces a deterministic output. That's the biggest thing I think we're missing right now, because With things like Jev. Jev released, uh, about a month ago, um, and that was a tool which— it's, it's not even an LLM, it's a classifier. It's a classifier model which takes in an input and it, it quickly makes a decision. So I made a, I made a model off of that, um, which had it play poker. It was fast enough that it could make decisions to play poker. You know, that shows you that it can take in large amounts of text and it can produce a deterministic output that's, that's hallucinates much less than an LLM. Do you think AI has a place in medtech today? Or like, because we talked about permissions and hallucination, a lot of our— the products we work on have some sort of AI component. What are your thoughts on the maturity of AI with medtech? I think absolutely. I think, um, it's actually one of the best uses of AI is saving people's lives. I think incorporating it in the medical space it's going to result in better access to good healthcare, better outcomes for patients, and it's going to kind of democratize the way that we treat, treat people. You see, in Utah, they just released— I think it was called NOLA— is an acne platform that analyzes your face, and it's legally allowed to give acne prescriptions to people by the— by Utah. They allowed it because there's such a shortage of dermatologists in Utah in specific counties, they said this is a good solution, um, to, you know, kind of fill that gap, to kind of allow people to self-prescribe themselves using AI. Stuff like that, other states are watching that closely, and I feel like we could see that definitely roll out for more types of medicine and kind of telemedicine. With— back to AI and medtech, I know a lot of the use cases I see with AI that come across our desk are engagements we have, or AI doing some sort of image enhancement for X-ray or MRI image. So it's looking for vascular disease or something that the human eye may not be able to detect. Or it's taking a reading like an ECG, and there's like fine granularity within like the heartbeat, as an example, that can be indicative of a, underlying condition that maybe a doctor can't pick up. Uh, so what other— besides those cases, what other cases do you think are, are valuable with AI and medtech? I think anything where it takes a doctor, um, a long time to make a decision, um, or a trained professional a long time to make a decision, and a model can be trained to observe exactly what a doctor or other trained professional would be looking for in those, um, those instances. And quickly infer what, what the outcome should be. Do you trust, uh, autonomous driving cars that have AI capabilities? I, I completely trust it. I, I trust Waymos, um, a lot. I, I've taken a Waymo in San Francisco and LA, and it works pretty well. I mean, I thought with, uh, with the hills and everything, it would have, you know, sped up when it's going downhill and everything, and I thought I was going to get in an accident, but it worked really, really well. It was actually really funny. We were going to a stop sign at the same time as a police officer, and the police officer waved us to go first, and then we kind of waited for the police officer to go. And then at one point, the police officer started going, and then the Waymo just went through the stop sign. So that's the issue with self-driving cars, is if it gets into some sort of accident, there's no way to contact whoever's in charge, basically. At a wedding in Missouri with some of my wife's family, and one of them has a Tesla. And of course, everybody has a Tesla, wants to show it off, and they have a new personality, and they talk about range anxiety and their Tesla 24/7. Uh, and Teslas are pretty much self-driving now. So of course we had to like do an experiment and show me the Tesla and all the cool features. And one of the things they tried to show me, which Wasn't true because Tesla, you know, if somebody's walking by, it'll pick up the signature of the person. They were trying to say if you go to cemetery— so I went to cemetery— that it can see the people that are deceased in the graves. The Tesla can pick them up. I'm like, this cannot be true because they're not moving, there's no heat signature. Um, and it wasn't true. So someone got out of the car and walked around and you know, and you could see like the remnants of like where their signature was, but it wasn't like the people in the graveyard. Have you heard— have you heard of that situation or that kind of stuff before? I have heard of that, yeah. I've, um, I mean, there's a lot of stuff with Tesla. Um, I mean, that's just a hallucination of its software. I mean, you look at the screen while you're driving any sort of Tesla, um, and you can see that it shows, you know, cars that aren't there, or It thinks that people are, you know, some object. It thinks it's like a car or something, or it'll hallucinate like a truck driving into you. You know, that kind of makes me scared because it looks like, you know, that's, that's the technology that the self-driving is based off of too, is the one that's showing the projection. When you're at a stop sign with a Tesla and it changes, the light changes to green, but you're in the turn lane, you're not in the, the lane to go straight. It'll beep as if the light turned green for the turn lane when it's talking about the cars going straight. So it'll tell you that it's good to turn if you're completely relying on it when it's not okay to turn. And, you know, that's the same technology that self-driving is based off of. So Tesla's technology is based on cameras only. You know, it would have cost a few thousand dollars extra to add LIDAR to these cars like the Waymo. Whereas Waymo. Yeah, exactly. Waymo's technology is a lot more accurate because they built it off of, you know, LiDAR, and they can afford it because they only have a certain amount of cars. Tesla's mass-producing cars, they can't just add in LiDAR because then no one would buy the cars because they're too expensive, um, unless they want the full self-driving functionality. Um, but I think we're going to see that Waymo's technology is going to come out ahead because they incorporated LiDAR, because the car has no way of knowing how far objects are away from it. which is a little bit concerning when you're actually in a Tesla. There's one instance of— I think the Wall Street Journal did a good investigative piece on it— of there was a crashed semi-truck on the freeway and a Tesla ran straight into it because it had never seen a crashed semi before. Whereas Waymo may have stopped because it's going based on distances to objects, so it might have realized there's an object ahead, I need to slow down. Whereas Tesla's just seeing Does this image match what I've previously seen? Um, yes, no, and then make decisions based off of that. I guess that is true, because I've ridden Teslas and I've taken a lot of Waymos, and on the screen, if you look at the Waymo, it is way more accurate. Like, it can show— it can find a little cat, a kitten, like, you know, in the corner of the, the hiding behind a car, as example. And the Tesla doesn't have nearly that much detail. Yeah, I agree. I think, um, over time we're gonna see, uh, whose technology is better, but I— my bet's on Waymo. And I think they're coming out with, uh, these— I guess like Waymo, but apparently like in Houston and a couple markets you can buy— you can basically buy a Waymo that just drives yourself around. Um, have you heard of that? I've heard of the CyberCab where you can buy a, uh, the CyberCab. That— yeah, that's what I'm talking about. I don't know if they've released that for sale yet. I think there was like pre-orders or something, but I don't know. But that, I mean, that would be interesting. I just don't know how that's allowed to be sold. But when you're in a Tesla full self-driving, you can't, you have to look at the road. I mean, there's like a specific sensor. Right, that's what I always wonder, yeah. Like if your eyes wander, it tells you, right? Exactly. And the issue is with that, people used to put weights on their steering wheel to make it look like, like the car sense that they're still driving, you know, quote unquote driving, um, and put a fake, um, you know, head looking at the, uh, or a photo of them or something looking into the camera. Yeah, people do that, then they just sleep while they're driving, which is extremely scary that people are just— but you know, if the technology is the same between the CyberCab and the Full Self-Driving, I don't understand why, you know, some people need to, uh, keep their eyes on the road and then the CyberCab has no oversight over that. I guess people are creative. I know a lot of people have been busted for the, the HOV lane. They would put like a mannequin next to them when they're driving to work so they could take the HOV lane and bypass all the traffic. I haven't seen that specifically happen, but I know someone who got an HOV violation. It was like $400. So avoiding that is definitely a good idea. Yeah, well, I thought in Tesla— in, um, California, because you're in California, If you have a Tesla, you can take the HOV lane just because it's an EV, right? If you look, every other car in California on the road is a Tesla. I'd say like pretty much 30% or so, 40%. So I don't think it's— I mean, I think they previously had done that. If you got some sort of sticker with the Highway Patrol, you could put it on your car and mark it as an HOV and then go through take the HOV lane. But I don't think that's, um, the case anymore because so many people have Teslas. I think they phased out the program, but I'm not completely sure on that. Cool. Well, we're coming up on time here, so what's, um, what's the takeaway from our discussion here? Like, what would you say a top takeaway or top 2 or 3 are for our listeners about, you know, medical device cybersecurity or AI? What do you think? I would say it's an exciting time to, to work in this field. I think that with medtech, there's a lot of, um, new opportunity that's going to come from AI with new things being possible that weren't before, especially with the increasing capacity of these large language models to take in vast amounts of information, you know, run for hours at a time. Um, you're seeing it, a little taste of what's going to be, um, the future with what, what it's done already, making good progress on the, um, on a lot of old math problems and with science. A lot of data processing is going to become a lot more streamlined and a lot more efficient, and we're only seeing the beginning of it. So I think it's a really exciting time to be working on something. As a reiteration of the excessive agency, I think that's how you put it, we need to be careful giving any AI agent or Claude or ChatGPT or anything we're using on our computer or accounts excessive agency, 'cause if something goes wrong, like you mentioned the example of it taking that guy's bank statements and sending it to his coworkers, it may have thought, oh, this is a good idea. That's a good example of excessive agency with AI. I believe we should take agency over our lives personally, but not let AI, you know, have excessive agency. So yeah, thanks for joining us today, Liam, and thanks for being part of Blue Goat, and, Uh, any, uh, last minute thoughts? I'm very excited to see what happens with everything AI related, and, um, I think it's a very exciting time to, to be working in, in cybersecurity. All right, awesome. Well, thanks everyone for tuning in to the Med Device Cyber Podcast. Hope you found some valuable, uh, information in this episode. I know we talked a little bit about Tesla and some of the Political items a little bit, but that's okay. We like to deviate a little bit here. So thanks, and I hope to see you on the next one.