
Patching Challenges in Modern Enterprises - BTS #86
Paul Asadoorian, Vlad Babkin, and Chase Snyder examine how AI-assisted vulnerability discovery is changing infrastructure security, why faster exploitation does not always mean greater risk, and what defenders can do when patching cannot keep pace.
AI is accelerating vulnerability discovery and exploit development, but finding a vulnerability is only one part of a successful attack. In this episode of Below the Surface, host Paul Asadoorian joins Vlad Babkin and Chase Snyder to examine the difference between technical exploitability and operational risk. From Linux kernel vulnerabilities and GPU rootkits to authentication bypasses in firewalls, the conversation explores why the growing volume of security disclosures demands better ways to determine which threats actually matter.
The discussion begins with malware that secretly mines cryptocurrency using compromised GPUs while manipulating NVIDIA monitoring tools to conceal its activity. That leads to a broader examination of the AI infrastructure attack surface, including management interfaces, software frameworks, drivers, and hardware. The team considers whether attackers could eventually aggregate stolen GPU capacity to run inference workloads or manipulate AI models without investing in expensive compute infrastructure.
Paul, Vlad, and Chase then challenge the assumption that faster exploitation necessarily translates into more successful attacks. They discuss reported exploitation windows as short as 35 minutes, AI-assisted discoveries of hundreds of vulnerabilities, and the practical constraints attackers face when turning an exploit into a sustained operation. For defenders, the more immediate problem is determining actual exposure amid an expanding volume of disclosures, particularly when vulnerable code may be unreachable or disabled by configuration.
The episode closes with a series of incidents involving fake Rodecaster software, compromised websites, remotely managed refrigeration systems, and industrial gateways. Together, these examples reinforce a recurring theme: patching remains essential, but security teams also need effective hardening, accurate exposure assessment, and the ability to detect unauthorized changes across the infrastructure they depend on.
Key Topics Covered
- GPU malware and hidden cryptocurrency mining: How attackers use NVIDIA management libraries to inspect GPU capacity, deploy Linux kernel rootkits, and manipulate
nvidia-smireporting so compromised systems appear to have little GPU activity. - The expanding AI infrastructure attack surface: Why frameworks such as Ray, GPU drivers, management tools, hardware, and firmware all create potential attack paths. The hosts revisit ShadowRay 2.0 and discuss how compromised compute systems can be repurposed.
- Stolen GPUs as an illicit compute cloud: Vlad and Chase explore the possibility that attackers could aggregate compromised consumer and enterprise GPUs for AI inference, model training, or password cracking instead of conventional cryptocurrency mining.
- AI model manipulation and weight-space programming: How changes to model weights could alter behavior without full retraining, and why hardware-level AI infrastructure attacks complicate the task of verifying that a running model is trustworthy.
- Faster exploitation versus operational capability: The hosts question the significance of reports that attackers exploit vulnerabilities within 35 minutes. Developing an exploit is different from building the infrastructure, access paths, and operational processes necessary to use it effectively.
- AI-assisted vulnerability discovery and technical debt: A discussion of OpenAI’s reported discovery of 948 vulnerabilities, primarily in the Linux kernel, and why larger numbers of discovered flaws do not necessarily translate into proportionally larger numbers of successful compromises.
- Configuration-dependent exposure and real exploitability: Paul describes analyzing an embedded device’s firmware to determine whether recent Apache and SSH vulnerabilities actually applied. The discussion connects to php2shell exposure in network appliances, where affected components and configurations influence practical risk.
- Prioritizing vulnerabilities beyond CVSS scores: Why authentication requirements, attack access methods, exposed services, and installed components can matter more than severity ratings alone. The hosts use firewall vulnerabilities and the growing volume of InfraTrust Pulse advisories to illustrate the problem.
- The risks facing network appliances and management interfaces: How firewalls, gateways, and other network devices can expose powerful administrative functions, allowing attackers to bypass access controls or expand their privileges after gaining an initial foothold.
- Hardening as a complement to detection: Why reducing available utilities, restricting privileges, tightening configurations, and monitoring changes can force attackers to take additional actions that defenders have opportunities to detect.
- Patching at enterprise scale: The challenge of updating tens of thousands of geographically distributed systems, including devices nearing end of support. The hosts examine maintenance windows, operational downtime, N-1 patching strategies, and compensating controls for systems that cannot be immediately updated.
- Fake Rodecaster updates and software supply chain trust: Unit 42’s discovery of Santa Stealer, distributed through fake Rodecaster software, illustrates why organizations need to validate software sources and cryptographically verify updates before trusting them.
- Compromised websites and malicious JavaScript: Research identifying 15,579 compromised websites serving malicious scripts, including payment skimmers and fake human-verification prompts, raises questions about coordinated vulnerability notification and remediation at scale.
- Refrigeration attacks and physical consequences: Reports of attacks on remotely managed refrigeration systems in French grocery stores illustrate how network-connected operational technology can create risks involving food spoilage, business disruption, and potentially sensitive medical storage environments.
- Industrial gateways and remote-management compromise: An authentication bypass affecting a Node-RED administrative interface in a Siemens industrial gateway shows how automation platforms can provide attackers with operating-system command execution and access to connected industrial environments.
Timestamps
The timestamps below are taken from the episode transcript.
00:00 – Episode introduction: GPU malware, hacked websites, firewalls, refrigerators, and industrial gateways
00:48 – Eclypsium resources and the opening GPU cryptomining story
03:08 – Management interfaces as unavoidable attack surfaces
03:39 – NVIDIA libraries, Ray, ShadowRay 2.0, and hidden GPU exploitation
05:32 – Could attackers use stolen GPU resources for AI compute instead of cryptocurrency mining?
08:10 – Model training, malicious model modifications, and compromised AI infrastructure
10:28 – Weight-space programming and changing model behavior without full retraining
11:29 – CrowdStrike’s 35-minute exploitation findings and AI-assisted offensive tools
12:53 – The operational infrastructure required to exploit vulnerabilities at scale
14:15 – Mandiant M-Trends, exploitation before disclosure, and the limits of speed metrics
15:48 – OpenAI’s vulnerability discoveries and the distinction between exploit availability and practical attack paths
18:31 – Linux MPTCP, SSH, and Apache vulnerabilities: testing actual exposure in device firmware
20:20 – What would make an AI-discovered vulnerability genuinely catastrophic?
21:16 – Technical debt, system diversity, and configuration-dependent exploitability
23:06 – The need for compensating controls and configuration-change monitoring beyond CVE matching
26:40 – Determining actual vulnerability exposure through components, configuration, and reachability
28:13 – How AI shifts the bottleneck in vulnerability management
29:28 – Zyxel and Barracuda firewall vulnerabilities, authentication bypasses, and prioritization
31:50 – Zero trust, credential abuse, and living-off-the-land techniques
33:12 – Hardening and detection as essential defenses alongside patching
35:12 – How hardening creates more opportunities to detect attackers
37:21 – Why patching remains critical despite rapidly increasing disclosure volumes
39:33 – N-1 patching, operational risk, and the potential for AI-assisted software testing
41:16 – Updating large device fleets and managing end-of-life hardware
43:20 – Geographically distributed critical infrastructure and remote-management risks
45:59 – Fake Rodecaster software, Santa Stealer, and verifying software and firmware updates
47:20 – Smart TVs, consumer electronics, and potential enterprise surveillance exposure
50:41 – 15,579 compromised websites, malicious JavaScript, and vulnerability notification challenges
52:54 – Refrigerator hacking in France and the risks of compromised physical infrastructure
55:17 – Remote refrigeration management and potential lateral movement
56:11 – Siemens industrial gateways, Node-RED, authentication bypasses, and command execution
57:27 – Closing thoughts on interconnected infrastructure risk
Links & References
Core References
- Eclypsium — Infrastructure security research and technology discussed throughout the episode.
- Eclypsium listener resources — Guides and research covering AI data center security, network edge devices, and infrastructure assurance, mentioned by Paul during the introduction.
- InfraTrust — Eclypsium’s infrastructure risk intelligence resource. Paul references InfraTrust Pulse as a way to understand vendor advisories and prioritize vulnerabilities.
- Below the Surface, Episode 85: Unveiling AI Infrastructure Vulnerabilities — The previous episode’s research into manipulating running AI models through underlying hardware infrastructure.
- Eclypsium and NVIDIA Collaborate to Address Emerging AI Infrastructure Attack Surface — Related Eclypsium research on AI infrastructure and model integrity.
Vulnerability and Threat Research
- Php2shell: A Network Edge Risk Playbook — Eclypsium analysis of PHP vulnerabilities affecting embedded network management interfaces, vendor patch availability, and end-of-support devices.
- Don’t Play with Fire: Prioritize Zyxel Firewall Update — Earlier Eclypsium research on remotely exploitable weaknesses in Zyxel firewalls and VPN appliances.
- Zyxel Firewall Vulnerabilities Reveal the Complexity of the IT Infrastructure Supply Chain — Paul’s analysis of firewall exploitation affecting Danish critical infrastructure.
- Fake Kernel Threads Are Hiding Backdoors on Linux Servers and Appliances — Related research on malware concealment, Linux processes, and the difficulty of identifying implants on infrastructure devices.
- CrowdStrike exploitation-speed research — The report Paul cites describing hackers exploiting flaws within 35 minutes.
- Mandiant M-Trends and Verizon Data Breach Investigations Report — Reports Chase references when discussing the changing time between vulnerability discovery, disclosure, and exploitation.
- Unit 42’s Santa Stealer research — Reporting on information-stealing malware delivered through fake Rodecaster software.
- ShadowRay 2.0 — An attack campaign targeting the Ray framework, discussed in connection with unauthorized GPU workloads.
- Compromised website JavaScript study — Research cited by Paul identifying 15,579 websites serving malicious scripts, with an estimated 135 million monthly visits.
- French refrigeration system attacks — Reporting discussed by Chase on remotely manipulated refrigeration units and the potential consequences for connected operational technology.
- Siemens industrial gateway / Node-RED vulnerability — Research discussed by Paul involving an administrative authentication bypass and operating-system command execution.
Tools, Platforms & Frameworks
- NVIDIA — GPU hardware and management tooling used in the malware story.
- NVIDIA System Management Interface (
nvidia-smi) — The GPU monitoring utility whose reporting attackers sought to manipulate. - Ray — An open-source distributed computing framework discussed as a target for unauthorized access to GPU workloads.
- Node-RED — A flow-based programming tool used in automation and industrial systems.
- Shodan — An internet-connected device discovery service discussed in relation to exposed infrastructure and vulnerability research.
- CISA — The U.S. Cybersecurity and Infrastructure Security Agency, discussed in the context of helping notify vulnerable website operators.
- UK National Cyber Security Centre — Referenced as a potential channel for coordinated notification of compromised websites.
- Weight-space programming — A technique discussed by Vlad in which model weights are modified directly rather than through conventional full-model retraining.
- Zero trust — An access-control model the hosts examine in the context of compromised devices, credential abuse, and unauthorized actions.
- N-1 patching — The practice of deploying the previous stable software version instead of the latest release to manage operational risk.
Transcript
Paul Asadoorian (00:04): This week, hacking the Rodecaster, websites are hacked, so are fridges, Zyxel firewalls, Barracuda firewalls, and industrial gateways. We have so much more. Stay tuned, below the surface, coming up next. Welcome to Below the Surface. It’s episode number 86, being recorded on November 13, 2026. I’m your host, Paul Asadoorian joined by Mr. Vlad Babkin. Vlad, welcome. Mr. Chase Snyder is here with us. Chase, welcome.
Chase Snyder (00:45): Greetings Paul. Greetings Vlad. Great to be here.
Paul Asadoorian (00:48): Yes. Quick announcement before we dig into it, below the surface listeners can learn more about Eclypsium and how we help organizations assure and secure the infrastructure their operations depend on by visiting eclypsium.com/go. You’ll find our guide to AI data center security, our white paper on eradicating hidden threats in network edge devices, and a case study showing how a major cloud provider used Eclypsium to protect its infrastructure hardware during a merger IT consolidation. You can also request a demo to see how Eclypsium verifies device integrity, hardens against risk, and detects active exploits across enterprise infrastructure from AI data center to network edge to user endpoints. Find it all at eclypsium.com/go. I want to actually start with the story that we didn’t even get a chance to put in the show notes that I thought was interesting, and we’ll go back and add it. And that is there’s malware running around out there. Compromising systems to do crypto mining, which has been a thing for a really long time. However, this particular malware is collecting telemetry from your system’s GPU resources. So it uses the NVIDIA management library to collect hey, how many GPUs are on the system? What’s their capacity? Sends that back to the attackers. The attackers then install a Linux kernel level root kit in addition to backdooring some libraries They conduct their crypto mining operation. All the meanwhile, if you query the system from user land or even as root, you it’s gonna tell you that your NVIDIA GPU’s barely being used. Except in the background, they’re mining cryptocurrency. Because that Linux kernel level rootkit, in addition to backdooring that library used by the nvidia-smi command, is actually lying to you. And so This is a I thought was a really actually a pretty slick way to mine some cryptocurrency. And it goes to show you that the motivations aren’t always, nation state level attacks. There’s still several groups out there just looking to mine some cryptocurrency. Which I didn’t even think was still a thing, but apparently it is. It’s still very commonplace apparently. Are you guys quiet ’cause you’re
Chase Snyder (03:08): Management interfaces are always dual use. That’s pretty much what it is. Any sort of management system is inherently it gives you the legitimate user capabilities that you want to manage your stuff, and it also creates an attack surface. And that I see no way to avoid that and whole vast industries of cybersecurity potential for trying to mitigate it because you can’t not have a management interface. And every management interface will be attacked.
Paul Asadoorian (03:39): It’s so there’s actually a lot of attack service here, right? So the NVIDIA management library is one. There’s also the Ray framework, which was targeted by ShadowRay 2.0 last November, which use very similar techniques. And they’re attacking the Ray framework, which is an open source Python framework for distributing jobs across GPUs. And it’s actually flaws in that particular open source application that gave them access to run code on the GPUs and then they’re trying to hide that activity through more attack surface in NVIDIA libraries and utilities that are used to report on that. Such as NVIDIA’s nvidia-smi, which is like a I think stands for system management interface. And that’s the utility that uses the libraries that reports on your GPU usage, for example. So as if we didn’t uncover enough AI attack surface, there’s a lot more out there. There’s libraries, there’s applications, there’s hardware, firmware, drivers, all of it is in play. And it kind of likens me back actually to working at the university where threat actors were back then targeting infrastructure largely for the resources, right? They would attack university networks, they would compromise large what back then were large servers, right? Like we had a fleet of Solaris servers back in the day. And we had super high capacity bandwidth that not many people had 25 years ago. And so the criminals, malicious actors, I don’t think we call them threat actors back then, would compromise that infrastructure and use it to serve malware, use it to serve pirated materials and whatever they were doing 25 years ago. A lot of what I said. Actually. Vlad.
Vlad Babkin (05:32): Yup, that makes sense. I’m surprised it’s not even used for like more compute, like… Right now, considering the scale of the internet and how powerful our personal devices, like especially those with GPUs, I’m surprised that they’re not using it for actual compute, like hash cracking or maybe even AI compute. So AI compute is a little bit harder to imagine in this setup. Like you actually need data center connectivity, which you don’t get with these systems. Crypto mining is probably the least creative use if you have a backdoor of that scale and that level.
Chase Snyder (06:08): I love this premise. You’re positing that they could basically create a neo-cloud out of stolen GPU compute. There’s marketplaces for the cheapest inference, and there’s whole companies, there’s data centers that have excess capacity, there’s marketplaces that essentially function as brokers for that capacity. This is a legitimate business. And you’re saying that someone could use this attack surface, this, you know. Break into residential slash individuals, gaming laptops that have a GPU on it, measure the excess GPU capacity, aggregate it, and then sell that for inference or use it for their own inference or training. I this is the most cyberpunk dystopian proposal that I have heard for as a use case for
Vlad Babkin (06:57): But just think about it. If you broke into a system with a 4090, de facto it’s a system with one 4090. And this can be used for inference on smaller models easily. Like you don’t even need connectivity. The only question is where do you hide a few gigabytes of data so that user does not… Hey, why the hell is my full disk? Because it’s now eaten.
Chase Snyder (07:21): They’re gonna hear the fan. They’re gonna be Why is my fan on this th device going off so much?
Paul Asadoorian (07:27): Yeah, Chase, sorry, could you just turn your gain down a little bit?
Chase Snyder (07:32): Sorry, I got too excited. Got too close to the mic. I was too hyped about this concept.
Vlad Babkin (07:32): Yes, it’s just…
Paul Asadoorian (07:32): We love hearing we love hearing you, but it’s just a it’s a scotch too high. Yeah, you’re clipping yeah. Clipping a little bit, it’s fine. It’s fine. It’s good. You get excited. Get excited, right? But yeah, you’re just clipping a little bit.
Vlad Babkin (07:36): The stupid part is… You don’t even need to hear the fan in many cases, like inferences, like if you don’t run it, stop, but run it occasionally or just a little bit. It’s not actually that bad on the load side. Yeah, like it will load your system, but not to a point where your fan will just spin. It will just be small spikes. Again, depends on how much you use it for, but that’s the stupid part. You don’t need to use it.
Paul Asadoorian (08:10): But I wanna go back I wanna go back to your point, Vlad, because I think it’s really good when you were going down a path of like if a threat actor has access to a GPU cluster or AI infrastructure, what would they use it for? Cryptocurrency is one thing, but I think using it to train a model is yet another thing because as we has Shodan our undermined research, that if you can influence a running model or replace even parts of a running model with your own training data, you still need to go train that model somewhere. And why not break into someone if you’re a threat actor, why not break into someone else’s AI infrastructure, train a model or it maliciously, and then that’s what you use when you infect other systems to insert that model. Microsoft actually just published some research where they were able to modify a model I think without retraining, but also their research is very kind of close, adjacent dotted line to ours as well. And I think this is going to be a growing trend as AI becomes even more pervasive in our computing environments, the threat actors are going to take advantage of it. And I think using our resources to train models, as hardware becomes very difficult to acquire, or they just don’t want to bother paying a cloud provider to do that for them, this could be something that maybe not your nation states are doing, but this is something that I think criminals are going to do that don’t have the resources and or budget to just buy hardware or rent the compute equipment required to train a model, which is very expensive, right? Nation states can like whatever. They mean We actually saw some reports that China does have some of the hardware that they’re not supposed to have. I saw an article about that today too. So it’s hard to control access to this compute, but it is hard to get. We all deal with it today. We I would run local models. Well it’s like good luck buying the RAM and GPUs to run that. It’s expensive. So criminals are just gonna break into other people’s systems and do it.
Vlad Babkin (10:28): Yeah, and to be honest, like what we have observed is even I don’t remember what article was it, but I believe somebody actually managed to use Claude to do brain surgery on Qwen or something like this. So like.
Paul Asadoorian (10:40): Really? When you said brain surgery do I thought you’re going to like actual brain surgery.
Vlad Babkin (10:45): I told it to brain surgery, but what they did was alter some weights to actually change what the model does, and they did it directly through inference with another model. And the whole point is, they called it weight-space programming, which is, to be honest, a pretty darn good term for this. So, it’s literally one model altering another during inference time a little bit, so you might not even need full training. We just need a bit of compute for this and not a full data center compromise to train one of the larger models. And the alteration might be very small, like where you will not notice it on most normal prompts and workloads up until the poisoned ones come up.
Paul Asadoorian (11:29): Well, attackers are moving fast today. We had an article from CrowdStrike that hackers are exploiting flaws within 35 minutes. I think that kind of dovetails off of our AI discussion as AI attackers gain more and more access to AI models, to AI assisted tools and frameworks such as Ardex. I don’t think I had that article here, right, but you probably saw in the news this week. Threat actors were using Artix. That one’s interesting because it is a pen testing framework, but the original author I think took it closed source, but there’s lots of forks and variations of it. However, those variations, before you start go looking for those variations ’cause you’re curious, a lot of those variations or forks are backdoored. Kind of like exploits are and were very much back in the day. If you downloaded a weaponized exploit, it would have a backdoor that hacked your system. So be cautious of that. But access to this technology is obviously I think fueling the hackers exploiting flaws at a rapid rate. I think we probably could have had four or five articles talking about this trend that attackers are moving faster. What have you guys been observing in this aspect of attackers just moving faster?
Vlad Babkin (12:53): So I’ll put it this way. Hey, Paul, here is a new vulnerability. Here is a POC by the authors. Make some news out of it. You have 35 minutes. Good luck. Like even with all of the access to cloud, to AI, if you don’t have an industrial grades post exploitation infrastructure where to exfiltrate this data, the best you can do in 35 minutes is probably CryptoMiner, right?
Paul Asadoorian (13:22): And you’re talking about but I love this, Vlad ’cause you’re talking about the full and it’s one thing to find a vulnerability, develop and exploit, and even have some post exploitation, but to conduct an operation from end to end you need a lot you need infrastructure and in a lot of other things.
Vlad Babkin (13:23): But like any serious exploit work You need capacity, need to like find the target to even exploit this ad. Okay, you can just find a bunch of random servers and show them this vulnerability. Great, you exploit them to what end? What are we gaining from this? All of this attention grabbing, no, attackers can weaponize something in 35 minutes. No, are know, apocalypse is coming. Guys, come on. The attackers who can weaponize this in 35 minutes are the ones who can weaponize it in… 24 hours. And for most intents and purposes, at this scale, 35 minutes to 24 hours does not really matter, right? They became slightly faster.
Chase Snyder (14:15): Yeah, totally. It’s a there’s a point of diminishing returns. We’re getting faster. For the past several years, all the big cyber research reports, the Mandiant trends and the Verizon data breach investigation reports and many others have reported on the mean time to exploit or this question of how quickly does a vulnerability get exploited. And for at least several years now, it’s been Many vulnerabilities get disclosed because they are exploited, and that’s how defenders find out about them. And I think the most recent Mandiant M-Trends report that I saw about it said it’s minus seven days. So vulnerabilities average on average get exploited seven days before they get disclosed and reported. I’m not sure if that was a subset of vulnerabilities or across all of the ones that they looked at for that M-Trends report. So yeah, as Vlad said. 35 minutes or 24 hours when the amount of time that it takes to patch a critical known exploited vulnerability can easily extend into days or weeks or months. It doesn’t really matter if they’re exploiting it in 35 minutes or 24 hours. They got all the time in the world, but it’s, a fun kind of hyperventilating story to put out there to say they can do it in 35 minutes, but it’s not making a meaningful operational difference to Really the attacker the defense.
Vlad Babkin (15:48): I will give you another data point. No, we just got a Linux release with 900 plus fixed vulnerabilities. Like something like this recently happened, right? And no, open AI.
Paul Asadoorian (15:57): Yeah, OpenAI open AI spent $1.2 million in the equivalent compute time and found 948 vulnerabilities, most in the Linux kernel and some in Firefox.
Vlad Babkin (16:02): Mm-hmm. They exploit all of them. The point is they exploited all of them, so they had an exploit for each single one. But the point is, for an attacker, doesn’t make a difference if you have one or you have 900. If they’re all patched in the same release, it’s better that they’re patched now. Okay, OpenAI can reproduce all of the exploits, attackers only need to reproduce one, right? So attackers with enough resources, the ones that you should be afraid of, even without deep AI assistance, They can probably find the lowest hanging one relatively quickly by just some quick filtering and like just research sense. We should develop once you work with a certain thing for a while. And like even if they have only basic AI systems and don’t have access to most advanced models, okay, it will slow them down by a day. You’re not gonna patch all your systems and that slow down one, two days. So the main scare is that, no, we will have a lot of exploitation. I have to do a range up on this. Where are the millions of ransomware sites? Like most major websites on the internet are not ransomware. Not by PHP to shell, not by the 900 Linux kernel exploits. Again, mostly because your average human will not be spending a million dollars with OpenAI to build all of this. It’s one thing having access.
Paul Asadoorian (17:33): Most of those most of those were most of those were privilege escalation exploits as well.
Vlad Babkin (17:39): Mm-hmm. This is a real second step. Point is to even get to a point where this exploit is useful for you. Even if you created this, let’s say somebody broke OpenAI’s access control and got access to their GPT unlocked for offensive use, right? Okay, they just exploited all 900 privilege escalations. They have all of the exploits lying around. Now what? If they don’t have a vector to even get onto the system, it gets them nowhere. And some of the exploits found no, PHP to Shell is going to end the internet and like everybody panic, everybody was patching, everybody was asking everybody else if they have PHP in their environments. A whole bunch of people were burning dinosaurs one after the other to get the exploit going. And as it turned out, successful exploit is incredibly hard. So yeah, I didn’t even produce the exploits.
Paul Asadoorian (18:31): It’s you know and it’s interesting. So I’ve been doing some spot checking on a lot of the recent vulnerabilities. So you’ve got a remotely exploitable Linux kernel exploit, the MPTCP. You’ve got the SSH hole vulnerability, and we’ve got one in Apache that was just released yesterday I think you pointed that out to me yesterday, Vlad, was the Apache HTTP/2 vulnerability. So
Vlad Babkin (18:55): Yup, there was one in Apache.
Paul Asadoorian (19:00): I was reverse engineering a particular firmware who will remain nameless. And that firmware had a Linux kernel and it was running Apache and it had SSH on there. And so I asked my AI harness, if you will. I said, You while you’re pull like you got this pulled apart, I’m let’s just do some preliminary checks. Is it vulnerable to X, Y, and Z? It was actually, nope, Apache does not have the HTTP/2 module. And by the way, it uses glibc and not musl, so it’s not vulnerable to SSH. And I don’t think I’ve yet checked it for the TCP one. But that’ll be my next question. So the first two I checked, they weren’t there. And this device is getting ready for our end of life. I think they’re still producing security patches for it, but I again I think Vlad you’re the optimist this week. In that the world is not burning that we do need to I think I put in one of my LinkedIn posts, we need to evaluate the situation and conduct the appropriate response. And we need to get comfortable with that process and iterate fast because things are just moving faster now. But it doesn’t mean everything’s vulnerable to everything all the time.
Vlad Babkin (20:20): Yeah, and like even then, I can give you one condition which will break my optimism immediately. And that condition is a world-ender type of bug. Think about the BMC bug that we disclosed before. But that is like as easy to exploit and isn’t everything. Or like something very popular, like say PHP. Because then it will be like everybody will reproduce it immediately using very simple models. And just go with it, right? Once such a bug exists and it is very easy to exploit, then we will have an actual apocalypse that everybody’s afraid of. But so far, with all of the effort to find vulnerabilities, in reality, AI just made us more secure. So far, it’s a net benefit. It’s not like those vulnerabilities didn’t exist. It’s just that AI eliminated them for everyone at once, instead of some attackers finding it and then hoarding it for years.
Paul Asadoorian (21:16): Well and that’s what people are saying and that’s what people are saying today is that AI is basically collecting everyone’s tech debt, right? Or maybe rea making people realize that you have that tech debt. Because it’s finding all of these vulnerabilities. And I think the one thing that may be saving us is that the internet and all of our services are largely are not like a monolith. And if you look at the really
Vlad Babkin (21:18): And using it in some offensive.
Paul Asadoorian (21:46): Super high profile vulnerabilities that came out recently, some of them I just mentioned. The internet isn’t burning. I mean unless you’re Microsoft and you had a day and a half outage, but that was unrelated. But if you look at how our systems are configured, they’re all different, right? You look at PHP, you look at your Linux kernel, you look at your Node.js, you look at your SSH, all as examples, not every system is vulnerable because there are variances in the configuration in the versions, in other kind of maybe not intentional compensating controls, but compensating controls just in the permissions, the configuration and the deployment options that you have for those technologies. And so that’s kind of saving us right now. Had this bug in SSH been in the core SSH code, it would have been much worse. I saw GreyNoise, I think, produced a thing where they were looking at like millions of scans for the SSH hold vulnerability and like 24 of them they can confirm were actually compromised. So that is helping, although if someone does find some more bugs that are, more ubiquitous, we might be in trouble.
Chase Snyder (23:06): Yeah, what you said, the diversity of configuration options and just the general non-monolithic nature of it reminded me of something that I thought of when we were talking about just the number of CVEs that are being discovered, but the relatively relative difficulty of generating exploits against them. But the direction that things are going strongly implies a growing need for some sort of non- CVE dependent compensating controls. Where if the new vulnerabilities are being discovered and published and talked about as if they’re the end of the world, and it’s very hard to tell which ones are going to be effectively exploited at scale or weaponized. And a ton of them are being exploited before they’re disclosed anyway. So by the time you know they’re actively exploited, you may already be. Owned by it. Some sort of compensating control that proactively measures for just changes, change monitoring at that configuration level. This is a little bit self-serving of me to talk about because this is something that Eclypsium does, but there’s a gap in the cybersecurity tool lineup for that kind of change monitoring and detection that. Does not need there to be a CVE or an advisory in order to flag potentially shady changes in the underlying OS or firmware or just the underlying systems in general or in the config of the kinds of devices that are getting targeted.
Paul Asadoorian (24:50): But there’s a difference between discovery and d and detection, right? Discovery is am I gonna be vulnerable to this? And you’re talking about this problem that we’ve talked about on the show before. That I think we have a an opportunity to really improve upon this in the industry and get away from that version checking and go, No, are you really vulnerable? Right? Now we’ve got this flood of vulnerabilities. What we need to do is automate the discovery. Like very good discovery. Do you guys hear that? I hear like what is that?
Vlad Babkin (25:26): Air conditioner guy, I will have to disappear for a bit.
Chase Snyder (25:28): I think that was Vlad’s doorbell and now he has to go talk to his air conditioning guy.
Paul Asadoorian (25:37): Or is that coming from one of my browsers or something?
Vlad Babkin (25:41): It was me, like I have, I will disappear for a bit.
Paul Asadoorian (25:44): I hear like l really loud music and Vlad’s muted.
Chase Snyder (25:48): Yeah, no I’m not hearing we’re not hearing music.
Paul Asadoorian (25:56): What is that?
Chase Snyder (26:05): You’ve been precision targeted to hear some loud jams. Somebody hit Paul with the bump up the jams vulnerability.
Paul Asadoorian (26:20): Where is that coming from? Sorry, I apologize to our live viewers. I’ll find it eventually. And now it just went away. That was really strange. Anyway, what were we talking about? Sorry, I got really greatly distracted.
Chase Snyder (26:40): Configuration, discovering whether you’re really vulnerable beyond just the presence or the association between a CVE and a version number being this very brittle mechanism versus reachability and other mechanisms for really measuring in a rigorous way your level of exposure or risk or vulnerability so that you can prioritize out of the glut of disclosures.
Paul Asadoorian (27:03): And we cause we talk about this in the InfoTrust Pulse, right? We’ve got hundreds of advisories, tens of thousands of CVE is released in a given month. Well, how do you shrink that down into more actionable intelligence? Well, I might be running the vulnerable version, but can I like knock out a whole bunch of those vulnerabilities by going, I don’t have the necessary components, configuration, or features enabled. So now you can whittle that list down even more. Then you can add your compensating controls. Like I’m less the likelihood is less that an attack will be successful because I’ve got some compensating control. Right? Sometimes those are built in, sometimes that’s user configurable, but I can r further reduce my risk. We need to rely on these things to help us prioritize. So when we get to work in the morning and there’s 2,000 new vulnerabilities, well I only need to know about or l care about one of those ’cause that one will have impact in my environment, right? I feel like there’s a ton of startups that have tried to help sift through this data. And it’s an ongoing battle.
Chase Snyder (28:13): Yeah, it’s all knobs. You they turn up the knob of risk or threat or exploitation against a particular area. Like as with everything in AI, it doesn’t solve an entire problem. It moves the bottlenecks around. And so the bottleneck to successful risk mitigation has moved due to the pace of CVE discovery and vulnerability disclosure, and to some degree exploit. Development and the bottleneck is different now and you have to change your security practice. And it also sort of changes the importance of different tool sets. It’s like there has been a long period in which vulnerability management and scanners were a really important part of the security stack. And I don’t think that they’re totally going away, but the level of at some point the overall volume And actionability of those. You need something else to make it useful. You know what I mean? Like the whole the landscape has completely changed which of your security tools are gonna make the biggest difference in your actual risk.
Paul Asadoorian (29:28): Yeah, and I think, it’s one of the things we also highlight in the InfoTrust Pulse is the class and access method for vulnerability matters more than the C V score. And we saw that this week, right? Both Barracuda and there was another fire is Zyxel. Zyxel a different beast altogether. I don’t know if that was an authentication bypass or unauthenticated out of bounds right to the into the switches, web management CGI enables OS command execution. I mean it bypasses authentication in a different way. Right? But those things that are vulnerabilities that are remotely accessible that do not require authentication tops of the list, right? Because that means if someone has access, your basic access control of who can manage this device is gone. And once that attacker has authenticated to the device, what I’ve found doing vulnerability research and looking at a lot of what AI is discovering. We just talked about those 948 vulnerabilities by open AI. Many of those are after someone successfully authenticates to the system. So we still have this very squishy center that we talk about, except it’s not so much on our perimeter as we would frame it before But it’s now it’s on all of our devices. All of our assets potentially have this soft and squishy center. So as Vlad was pointing out earlier, the attacker just needs one way in. Once they’re even just authenticated to your firewall as a regular user, their path to gain additional privileges and level of access like command execution, there’s far less barriers to that. And I think if you’re doing vulnerability research and bug hunting today probably having a grand old time with AI, hey, once I’m logged in, I can do co OS command injection authenticated, right? That’s a much easier flaw to find. But the bypassing authentication while much harder, way more important for enterprises to focus on today.
Chase Snyder (31:50): I would guess that organizations that have been pursuing zero trust as a model for how they secure their operations, really verifying whether a particular user or device has the privileges or should have the privileges to take a particular action are gonna struggle a lot to defend the whether their zero trust setup is really effective anymore.
Paul Asadoorian (32:24): Yeah. ‘Cause they’re like you’re it’s a great point, Chase. Effectively attackers are more and more living off the land. Hey, I can authenticate to a device, I can issue commands. And the favorite one that we’ve talked about, right? Is I’m just gonna passively listen for more credentials. And then I’m gonna log into other stuff and I’m just gonna do things with that other stuff that doesn’t jump outside the bounds and set off bells and whistles that something is not all even in a zero trust model isn’t authorized.
Chase Snyder (32:56): The potential to at the device level and even the hardware level sort of subvert or route around any sort of a zero trust like authentication that’s happening is just is growing a lot.
Paul Asadoorian (33:12): For sure. For sure. And I think it comes down to, detection. Prevention with patching is gonna be hard as we outlined. So you’ve gotta do that detection phase, and you’ve gotta do a lot of hardening. I think that’s the w only way we survive, right? Is being really good at detecting when something’s happening. And being really good, your prevention shifts, I think, from patching to hardening. And we don’t harden as much as we should. We don’t inspect our configurations, right? And that has to change, right? Because you’re
Chase Snyder (33:49): Yeah, I think that’s gonna change. The organizations that survive are gonna do hardening a lot be the ones that proactively go down the hardening path a lot faster.
Paul Asadoorian (33:57): Yeah. 100%. They’re gonna be the ones that are if you can get to a point and I think, when I used to do CTFs, this was certainly the case. Great. You’ve just gained access to my system. Okay, now what? Like good luck, because I’ve got it hardened down to make it super difficult for the attacker to do anything. I’ve removed util we used to get on Linux systems but there was like no utilities or whatever, right? Like there’s ways to really harden your devices. That I still think the game is if you make the attacker try harder, whether they’re using AI or not, that’s gonna become easier to detect. So they play off of each other, right? If I make my system harder, hardened, harder to navigate around and do anything with, the attacker’s gonna have to try a bunch of different things. Those are gonna set off bells and whistles. Whether it’s an attacker or an AI harness doesn’t matter. In fact the AI pen testing harnesses we talked about are very noisy today. Which is good for us as defenders, right? Until they learn to get more stealthy. But I think that’s really the strategy today because you’re just you the this identify and patch strategy is not sustainable. It’s not gonna work in today’s age.
Chase Snyder (35:12): The thing that okay, it’s been an it’s an old saw, right, in cybersecurity that the attacker only has to succeed once and you have to succeed every time. But I think what you just said kinda inverts that a little bit, where if you do the right hardening, you make it so that the attacker has to try a lot of other stuff. They have to knock on different doors, they have to Just generally poke at the system more to find a way. And that creates a lot more opportunities for you to discover them. And so it’s like the attacker, the defender actually has a lot of opportunities for discovery that are not centered around, matching the CVE to the version or whatever. But you have to, if you do that hardening, a secondary benefit that you get, other than just making it harder for them to attack you in the first place, is you create more detection opportunities. You create more opportunities where they’re gonna tweak something that you are watching for. And that I think is yeah, that’s a direction. That’s a mode of thought that I think will help a lot of organizations stay afloat while we’re in this kind of transition period of there was some article that said that AI is making us pay off 40 years of technical debt. We’re in this period right now where so many. Ancient vulnerabilities in deep foundational infrastructure code, like the EDK, reference code are being discovered, and that’s not gonna last forever, you know. There will be a point at which all the low-hanging fruit of the last entire history of software development has been picked and dealt with, but we’re in that transitionary period, right transitional period right now, and I think that organizations that do this hardening. And force the attacker to try more stuff against their system and then also build up the monitoring, the change monitoring discovery capacity where they’ll detect the other stuff that the attacker is trying. I think that’s a solid mindset to get through this period that we’re in right now.
Paul Asadoorian (37:21): I do also think like I don’t want to discount patching here. Maybe it was a little too hard on patching. Patching is a great lever to pull to reduce your exposure and your technical debt, right? Like the whoever’s providing that software is reducing the technical debt in your software. That’s translating to you when you implement that being on the later versions. Those later versions not only are squashing vulnerabilities, but also hardening as I go back to Fortinet. The later releases We’re mitigating those vulnerabilities. So it you still need to be doing this patching cycle, but also keeping in mind that you need to supplement that with something else. So don’t lose sight of patching because it does as you said, all these AI models like you see all these big companies, let’s just look at network edge devices. You see all these big companies going, We just took a front they actually come out and say it. We took a frontier model, we ran it against our code, and we fixed a bunch of vulnerabilities. Here’s a new release. You gotta start moving more quickly on that patching side ’cause that’s the known stuff.
Chase Snyder (38:26): Yeah, absolutely. It’s just that it’s patching that stuff is table stakes. And it’s hard. The capacity to roll that number of patches out is not there at a lot of organizations. When your network vendors puts out 10 times as many patches this month as last month, do you have the people? Can you absorb the downtime and the operational risk that it requires to roll those patches out? And so there’s this sort of gap where
Paul Asadoorian (38:31): Mm. It is. Mm-hmm. And upright risk.
Chase Snyder (38:53): Okay, you’re gonna roll all the patches out. That’s table stakes. If you’re not doing that, you’re already done for. But there’s a gap where it’s okay, our capacity to patch is this much, which has been, enough so far. And now the amount of patching we need to do has gone way up, and we haven’t been able to ramp up the capacity yet. And so the time it takes for us to get to those additional patches is gonna be longer. What do we have in place now that can help us stay secure, stay hardened, discover attempts, protect?
Paul Asadoorian (39:17): In the meantime.
Chase Snyder (39:22): Protect ourselves in the meantime while we ramp up to being able to achieve the new normal, the new greatly increased sort of baseline of patching that we have to do that is absolute table stakes for security.
Paul Asadoorian (39:33): And I get the and I get the N-1 patching strategy that you never want to be on the latest and greatest release because that release could have bugs, not just vulnerabilities, but bugs in general that introduce operational risk. And so it’s historically safer to do not the latest, right? But the version before that. Because other people have used it’s been battle tested. They’ve fleshed out maybe a couple more bugs with that, or maybe those bugs are in the next release. Which a case could be made either way. But if you’re N-1, you’re right, you need those stopgap measures. And I also get the operational risk is important. What I hope is that we’re able to use AI technology to have more reliable and efficient software so that customers can trust. I want to be on the latest release. Like I know the latest release has a less likelihood of introducing operational risk because the vendor has used AI not just to find vulnerabilities but to also test it. Right? AI is great for testing. Some may argue with me, because if it’ll hallucinate the right answer in your testing process, but I think it’s evolving and getting better at that. And I but so my hope is that we can get to a point where applying the latest software release, firmware release is operationally acceptable. It doesn’t incur that risk. I hope we can get there. ‘Cause we need to be. ‘Cause this onslaught is not gonna stop.
Chase Snyder (41:09): Fingers crossed. Some organizations are gonna get there and they’re gonna win the future, and some aren’t and they won’t.
Paul Asadoorian (41:16): Hm. Yeah, and I mean we can sit here on our pulpit, right? But let’s say you’ve got 50,000 of XYZ device deployed, and you’re not patching that overnight. And maybe you are. Maybe you are. But that’s a challenge, right? I think we sometimes think of it I as in related to our own personal, I have one Rodecaster as an example and I need to apply patches to it. Okay, that’s fine. But what if you’ve got 50,000 of some device deployed out to your enterprise and you need to patch all of them? You’ve gotta do that in stages, batches, you’ve gotta test that. Test your config against it, the whole thing.
Chase Snyder (41:57): Sure. And every organization, especially at a certain scale of enterprise, and more so in some industries than others, you’re gonna have stuff that’s end of sale, end of support, end of life, that’s still operating. And I find this to be more true in some of the critical infrastructure verticals. And I think that’s probably caused by multiple factors. Budget, operational downtime being more and more risky in certain critical infrastructure sectors. And so their willingness to shut stuff down to update it is reduced. And so there’s you’ve got 50,000 devices and you have to patch them, but they all can be patched. You just have to do it in segments and you’ve got to manage the operational risk. And then there’s organizations where it’s like you’ve got 50,000 devices and 10,000 of them are not going to get the patch. The vendor’s not out. And so you have to either do a hardware refresh at that point, which is a whole other level of operational difficulty, or you have to figure out some sort of a compensating control that reduces the risk of not patching to an acceptable level for you. And yeah, the logistics of that just get harder and harder as you get to that sort of end of sale, end of support.
Paul Asadoorian (43:20): You bring up great you bring up great points, Chase, about the challenges we have in keeping our infrastructure up to date. And I think sometimes we in cybersecurity we hey, you should just keep all your stuff up to date all the time, forgetting that those 50,000 devices may be geographically dispersed across the globe. And that presents problems. When an upgrade goes wrong, someone might need boots on the ground. Some of the a lot of devices don’t have backups. It’s just not it’s cost prohibitive to have a fully redundant infrastructure in every aspect of your business. So there’s a lot of challenges to keeping your infrastructure up to date, which I think more is more fuel to the fire for getting really good at the detection, the discoverability, in the hardening aspect of it. Hardening can also impact operationally those devices as well. That was like my first foray into security. I was I need to harden these servers and I’m Wow, I made that change and now they don’t work. I need to back that out. Go figure. Go figure. Still very true today.
Chase Snyder (44:33): Yeah, they’re complicated Rube Goldberg machines. Yeah, the geographic distribution is a really tough one. If you look at critical infrastructure industries like oil and gas, where, there’s a bunch of just like pump stations, there’s offshore rigs, all of these have IT infrastructure in them that is heavily resistant to being shut down at all ever for any reason. They have tons of compensating controls in place already to reduce the potential of being targeted and they want to be air gapped. But when being air gapped means you got to helicopter someone out there for any sort of IT update, there’s a strong incentive in place to figure out a way to be able to remotely do that. And then we get back to the question of sort of any management interface becomes also an attack surface. Where if they open up any sort of remote connectivity so that they don’t have to send a chopper, that remote connectivity has to be incredibly secure or it’s going to be targeted. And this yeah, this is a challenge that many of the major critical infrastructure industries face is the geographic distribution of mission critical systems that gotta either get updated or heavily insulated against these attacks.
Paul Asadoorian (45:59): I mentioned Rodecaster earlier. I did that on purpose because Unit 42 found some fake Rodecaster software. I happen to have a Rodecaster device. I’m actually not using it for this show today. But it’s a great device. However, it underscores the point of validating and verifying where your firmware and software comes from before you install it. This one was actually called the Santa Stealer. And if you installed this fake update for your Rodecaster. It can steal your browser passwords, cookies, autofill data, crypto wallets, Discord sessions, take screenshots, and fetch another upload. So this, definitely and this is a consumer personal device, but it extends to all firmware and software. Being able to cryptographically verify this software and firmware updates is hugely important. I mean it’s hugely important to our mission here at Eclypsium. And all vendors and enterprises need to have this, they need to push for it, because we need to validate this. If any someone’s able to distribute that fake firmware update, they will they can hide and own your system. This didn’t do a very good job of hiding, I don’t think, but certainly things can hide.
Chase Snyder (47:20): Yeah. I’m gonna admit I didn’t know what a Rodecaster was. So when you put that in the on the docket for today, I had to go look it up. And it’s like a podcast mixing, it’s like an audio professional audio tool, right? But yeah, you plug it into your computer, right? This reminds me of the recent story that went around. This was not on our docket, but speaking of consumer devices that get connected to important corporate systems and leave the way in. The Someone reverse engineered, I won’t say the brand, but a certain brand of television that they had in their home, smart TV that was connected to the internet and found that it had an always-on microphone that was listening to everything that they said and using that essentially to sell ad targeting data. But it’s easy to imagine a corporate office that has big TVs in the conference rooms that are connected to. Enable, remote conferencing inside of a conference room that also have this, that are have these, backdoor phone home capabilities. It’s quite it’s not a leap at all to see these seemingly consumer sort of gadgets turn into an entry point or a point of data exfiltration from yeah enterprises, governments.
Paul Asadoorian (48:40): It’s super hard even in an enterprise setting it’s still difficult to get screens and TVs that aren’t don’t have the smart capability. In fact the model that I had in the podcast studio was still a Samsung Smart TV and it still had Wi Fi connectivity in the whole thing. And these were supposed to be, a model that was meant for commercial use. But Samsung still would not remove the smart functionality on them. So you’re right, you could have the and it there’s been multiple report research reports on these TVs that have them doing all kinds of nefarious things, which is super interesting, and very dangerous for both consumers and enterprises having these in their environments. They’re just doing all kinds of shady stuff.
Chase Snyder (49:38): One of the big AI labs reported an incident recently where they had found an employee had placed a recording device behind a TV in a conference room and was trying to like record and take data out of this conference room where big important calls happen and stuff. And after the TV reverse engineering story came out, it felt very ironic to me that he had placed this separate recording device behind the TV. Cause I’m bro, you could’ve there’s already a TV in there. You could have just used. Could have just used that. This is obviously a joke, a silly, a silly premise. But the idea of the TVs being just connected and being phoning home is not a joke at all. That is absolutely happening, and I would guess that companies have those in their conference rooms, phoning home audio data and whatever it hears, in a way that the company would prefer not to have happen if they looked into what was actually being exfiltrated.
Paul Asadoorian (50:41): I mean everything’s getting backdoored, including 15,579 websites that were found serving malicious JavaScript. They estimate more than 135 million visits per month. The embedded scripts included payment skimmers, fake prove your human prompts that deliver malware, device exploits and redirects. They conducted a study, found these. The hard part Now is contacting the owners of these sites because there’s so many. And I think we deal with this when we look at any kind of research we do here at Eclypsium or other people’s research. Hey, there’s a vulnerability in this thing. Well, how many are on the internet? Right, is usually one of the first questions. And it’s Well, you can go to Shodan and you can identify that. And then I’m but who’s contacting the people who are vulnerable? Whose job is that?
Chase Snyder (51:34): Is there events for fifteen thousand websites to get reported to the
Paul Asadoorian (51:38): Yeah, they are so it does it’s funny you say that. So they are in touch with CISA, and NCSC says that route looks could look more promising, right? So you have to work with another agency to help just discover and then try and contact these sites. I think it’s one thing that we don’t talk about or maybe do enough is there needs to be like a there I think there was a couple of non profits spun up over time that look at how do we just have an automated process that contacts people who are vulnerable that just may not know it, right? They’re not listening to our podcast. They need to start listening.
Chase Snyder (52:20): Missed that detail from that story that some of what some of the malware was being delivered through fake prove you’re a human prompt. Talk about adding insult to injury. I hate doing those captchas or whatever. To do one of those and then get malware because I did it. That it makes me laugh, but it also kind of boils my blood. I the st the state of things right now.
Paul Asadoorian (52:29): Right. There’s some of those captions are terrible. Right. Chase, you had a story about the f the fridges. This is not a an old story either. There’s this is the second w I counted two waves of fridge hacking.
Chase Snyder (52:54): Yeah, they’re keeping fridge hacks and they have all the same sort of characteristics that we talk about. It’s like they’re not up to date. Who’s updating their fridge firmware? Nobody. Even if you’re a grocery store, or the more like the sort of thread that they drew is okay, all these grocery stores in France had their refrigerators hacked and the temperature turned up, like the thermostat in it, or some humidity controller or something like that was able to be accessed and they could influence it to turn up the
Paul Asadoorian (53:03): Apparently no one.
Chase Snyder (53:23): Temperature, which of course spoils all the food. It can do serious damage. I don’t know if there’s a health angle to it where it’s they were for a little while before they figured out what was happening, they were selling food that hadn’t been properly refrigerated or something like that. But then someone made the leap to this same gear, this same refrigeration gear could also be in use in medical situations where they store like blood or plasma donations or vaccines that need to be refrigerated. And that starts to, be a little scarier. It’s really bad for these businesses that probably had to throw away just millions of dollars worth of food. But yeah, it’s infrastructure-esque sort of attack. You hear you immediately hear about it, and there’s a lot of like fear-mongering when something like a water treatment facility or a power plant gets hacked. But the implications of these, yeah, refrigeration or this type of device, like it’s one step away from being a medical device or something like that. It’s it can have a serious public health impact if it’s pointed at the devices in the right contexts. And I don’t know that healthcare organizations have their refrigeration units better insulated from attack than grocery stores do. It’s possible. I just don’t know.
Paul Asadoorian (54:45): I like your usage of words there better insulated.
Chase Snyder (54:48): Ha. Unintentional, pun intentional.
Paul Asadoorian (54:52): Intentional. Yeah, and I don’t know who’s conducting these attacks. There’s a lot of speculation. Is it a test from threat actors? Is it shenanigans as I like to deem it? We just don’t I d I haven’t seen much about the motive as to why you would want to do this. Because it’s there, kind of thing. I d I don’t know. I really just don’t know.
Chase Snyder (55:17): The article, there was numerous articles about it. That and some of them hinted at this issue. A, that these are administered from a remote connection. So it’s like there’s tons of these refrigeration units out, and they don’t want someone to have to go, they don’t want to call up a fridge technician to go up to every individual fridge. So they have remote monitoring, remote management for hundreds or thousands of the units probably. So yeah, management interface. And also that they are network connected, obviously, to be able to be remotely managed in that way. And are they segmented from other critical systems inside of the, corporate IT of this grocery chain? I don’t know. The possibility of lateral movement from a refrigerator into more internal IT infrastructure seems pretty real. Pretty real to me. And
Paul Asadoorian (56:11): Yeah, and kind of an adjacent story too. There was a node red admin interface. So apparently Node-RED is like a the summary was it NodeRed makes it easy to build automation by connecting visual flows and some nodes can run operating system commands. So they found a an authentication bypass in the node red admin interface that allowed them to and this was is used in industrial environments to manage and monitor the connected industrial equipment, but attackers took over the management interface and were able to run commands on them. So kind of similar to the fridge thing too, right? It’s all part of what I deem as critical infrastructure or industrial infrastructure that can be very fragile and attackers are going after it, right? I’m not sure which attackers. This was a Siemens happened to be a Siemens product industrial gateway. That was targeted. But there’s a link in the show notes to the article. So Very scary stuff. Anything else, Chase, you want to talk about this week?
Chase Snyder (57:27): Man, I think we covered it. This was a zesty we had a lot of overlapping and interconnected Rube Goldberg machine of show notes as we have the Rube Goldberg machine of connected infrastructure technologies.
Paul Asadoorian (57:45): Outstanding. Well, Chase, thank you very much. I know Vlad could not rejoin us, but thank you, Vlad. Thank you, everyone, for listening and watching this edition of Below the Surface. We’ll see you next time.