In Episode 7 of Censys ARC Flash, Silas Cutler, Principal Security Researcher, and Himaja Motheram, Senior Security Researcher, test the “Vulnpocalypse” narrative against what Censys data shows for PHP2Shell patching, break down the open-source agentic pentesting framework ARTEX, and cover the second wave of attacks on Danfoss and Copeland refrigeration controllers across Europe.
Silas Cutler: All right, let’s kick it off.
Himaja Motheram: Howdy.
Silas Cutler: Awesome. Thanks for coming today. This is episode seven for the Censys ARC Flash series. It is November and we have a whole bunch of topics to cover today. Starting off with a couple of quick news stories that are going on. Europol just published another report as part of Operation Endgame. This time their operation was heavily focused on StealC, Gootloader, CastleRAT were also included as well.
Over to Himaja for the Danfoss updates.
Himaja Motheram: Yeah. So basically the hot news story of the past week or really it feels like the month has been these fridge hacks that have really taken Europe by storm. And so in the first week of November what we saw was the same sort of fridge controllers manufactured by Danfoss, along with some manufactured by Copeland, were targeted in a second wave of exploitation across more countries. So before they were sort of localized in France for the ones that happened in the first wave in October. And we saw them kind of spread across multiple countries this time, as well as a notable incident where a fridge was actually tampered with that belonged to a French blood bank ministry, which is kind of an interesting new angle where, the first attacks we saw mostly focused on retail and grocery and most of the losses were your chicken goes bad, but then we’re also seeing it start to lead into other parts of the cold chain. You have vaccine storage, you have blood storage. Luckily in the case of the blood bank they were able to see that the monitoring was there. They checked the monitoring and were alerted to the fact that it was happening and were able to save the blood.
But it just raises questions about whether refrigerators are used in all sorts of industries and are going to be a new area that could be considered critical infrastructure when it comes to operational technology.
Silas Cutler: I’d absolutely say so. And I think we saw the beginning part where it was supermarkets and that was a very impactful thing for average folks.
Himaja Motheram: Yeah. What we have seen actually when we look at the different types of fridge controllers that are exposed on the Internet in Censys, you can see in this graph that the first wave happened in October and the numbers for each of these devices covered around three thousand each, three thousand to four thousand. And since that first wave, the number of Danfoss, which was the one affecting the first wave, went down and then after the second wave, both of them have dipped a little bit and continue to dip past this. So that’s good news. But at the same time we don’t know it’s a great sign. But we also don’t know why they go offline. Are they being, you know, compromised or are they being remediated? But it is eventually a good sign.
Silas Cutler: Awesome. Well, let’s dive into the two major sections we have today. So we’re going to talk about the blog that you put out on the Vulnpocalypse and then at the end we’re gonna move over into talking about ARTEX a little bit. So let’s dive right in. So the blog that you published was “Thoughts on Vulnerability Response in the Era of the Vulnpocalypse and PHP2Shell.” Can you take us through some of the things that you talk about through the blog for people that may not have read it?
Himaja Motheram: Yes. So this title kind of gives it all away. But for those who may have been living under a rock, the Vulnpocalypse is a term that is being thrown around to refer to this idea that with the rise of AI and other agentic tools being used for malicious purposes and also for, you know, vulnerability research, that we’re just going to see this absolute storm of new vulnerabilities rain down on the security community and it’s just going to be a situation where all we know about how to use to defend ourselves against vulnerabilities are going to be moot because there’s just so much volume. So kind of a doomsday scenario, which the industry loves to do, and so do I every once in a while, but what this blog goes into is it kind of does two things. It looks at the actual claims being made, which are that AI is going to make traditional vulnerability responses obsolete, and we’re going to have to come up with new things. And compares that narrative to what we saw actually happen in the case of a recent big vulnerability that was discovered using AI, which is PHP2Shell. And we look at how in Censys data the actual patch response played out. Like how did the Internet respond to this vulnerability?
The overarching idea is that, the Vulnpocalypse, what we see is that, there’s so many that that more vulnerabilities and a higher volume of CVEs has never meant meaningfully more risk. Because a vulnerability does not necessarily mean that it will be able to be exploited, right? And what we see is that the volume of CVEs is going up, but the level of mass exploitation is not, we have a lot of claims of that, but the actual activity is not going up at the same time. And in the case of PHP2Shell, when we look at which orgs were patching and which ones weren’t, a lot of what seems evident is that the ability to respond to that was decided for an org way before the incident even happened. It was about, you know, were they positioned to find their assets? Were they positioned to mitigate them in a lot of cases, were they positioned to be able to easily apply a patch, you know? And so I think that there’s a lot that can be a lot of thud right now. And I wanted to talk about what is actually changing.
Silas Cutler: And I think you did a really good job with it. You called out a really interesting statistic in this, and that even with this mass that we’re seeing, it only ends up only being 0.023% of these vulnerabilities that actually turn into things that end up being mass exploited in the wild. And after a year, where Mythos seems like ancient history now, but Mythos came out, I think in April with the Glasswing project, and there was like a lot of this really began. But it seems like that while there are more vulnerabilities, the ability to actually exploit them is still difficult. Do you see that under one percent of mass exploitation changing at all in the future or or how do you think it’s gonna play out over the next year?
Himaja Motheram: I think that it’s really gonna depend on the vendor. So that 0.023% that number actually came from GTIG and it was specifically about Linux kernel vulnerabilities, which are sort of the the case study for how much, you know, how much vibe coded bug bounty and exploitation has has been happening because I think that they saw a surge of like some huge number of Linux kernel vulnerabilities. But then yeah, so few of them were actually observed being exploited and Linux, that’s an example of like, you know, an open source, like really popular operating system. And so PHP was an example of, again, a very popular open source tool that is embedded in all these other things, comes bundled with all these other software technologies. So I think what we’re gonna see is that those product those those types of vendors and those and like the types of software that do not have big security budgets behind them that are maintained mostly by volunteers are gonna get more and more heat, especially because, the impact, they’re not being maintained and they don’t have as much of like the the support behind them to to respond to these things. And then also they’re so popularly used. It’s kind of, you know, the relevant XKCD of the tiny thing that is being, yeah, the Internet is standing upon it. But I do think that like big vendors, we’re not really seeing it in, you know, like the big software yet. And maybe I’ll bite my words about that. But I think it is really what I’m seeing is these like small to mid sized vendors and projects that are being exploited the most.
Silas Cutler: It’s kinda like the shift from ransomware where it shifted going after the small to medium businesses because they were still able to get similar results to going after the large ones.
Himaja Motheram: Yeah. Yeah. Big fish.
Silas Cutler: It’ll be interesting to see how it goes over the next year. I do wonder if we’ve hit that inflection point where there’s been sufficient AI testing for a lot of these projects, but it’ll be interesting to see how it plays out.
Himaja Motheram: Yeah, and that’s another point I make in the blog is we’re talking a lot about how AI is changing vulnerability exploitation and weaponization, but I would love to see what is being done on the other side. I think that we’re getting to a place where it is mostly being used to do security checks, like you know, in CI/CD, or being used to find bugs in existing code. But I’d love to start to see agentic tools being used to actually fix like the structural, you know, proactive like design things that make a product insecure. I think that will go a long way for defenders.
Silas Cutler: Yeah. I mean of course hopefully it doesn’t introduce its own new categories of bugs. Like there’s some things that I’ve been tracking on like standard endpoints that now start to become much more common that that normally ought not be exposed. So it’s a brave new world.
Himaja Motheram: It is a brave new world.
Silas Cutler: So with that, let’s shift gears and we can talk about ARTEX a little bit and how that has also been a fairly hot topic over the past few weeks as well. So for folks that are just chiming in, We put out a blog post last week that talks about an agentic pentesting framework called ARTEX. It is also open source, but it’s also worth noting that ARTEX along with ScopeSentry, which is a related tool, and CyberStrikeAI, all three of which are actually developed by a security company, ThreatBook, as part of some of their commercial offerings. We’ve started tracking it fairly closely, but it definitely is gonna be one of these things I think folks are watching for at least the next month or so, just because the volume of it seems to have spiked up pretty high as well.
Himaja Motheram: Do you mean just the number of exposed hosts that we see or just like reports about it?
Silas Cutler: Both in this case, CrowdStrike and Microsoft both put out reports in October about the usage of ARTEX. CrowdStrike’s being focused on a South Korean financial entity which was targeted and Microsoft’s focused on Storm-0940. So it has been a really interesting thing to watch how attackers have been adapting it, because while it’s not necessarily creating new techniques, it distinctly looks like some of that traditional attack cycle phases, reconnaissance, exploit, instead of having that be a sequenced process, some of that work can now be done parallel. And some of it likely without even the actual attacker having a deep understanding of how those things work. They simply just provide a task that’s we’re going after this, use these external APIs. Let’s go!
Himaja Motheram: Yeah, it’s designed to run completely without human intervention, which is really really frightening. I was seeing some studies where it just continues, it can do so much without bypassing rules. I know that it was on GitHub in July of this year, and you mentioned that the exploitation and the reports came out in October, around October is also when the author took it off of GitHub. Yes, but as you mentioned, we’re still seeing these forks and rebrands of it like Scope Weaver and ScopeSentry and these other things. Do you think that taking it offline actually did anything or are we just gonna continue to see it proliferate?
Silas Cutler: I think we’re gonna see more proliferation. I do have some curiosities about how the removal of it is going to impact usage going forward. It might have been last Friday, there was a post from MalwareTech circulating, where it was a screenshot of the source code for the login function was placed side by side with a version that had a backdoor added that would execute commands through the CAPTCHA input field. So it looks like people are tampering also with some of the versions. So there is a mixture of legitimate forks that are from that original code base and forks that are from people trying to add malicious additions to it. I do wonder if that’s going to reduce the overall usage. I haven’t seen any that have that CAPTCHA field in our data, but also I don’t think we see that necessarily from the login page. But, it’s still something I’m very curious to see how that impacts adoption going forward. But we still see it continuing to rise in usage.
Himaja Motheram: Yeah. And you mentioned in the blog that the kind of the number we found in Censys, over fourteen hundred of them exposed, is probably an undercount actually, because it does not need to be exposed to the web to be used. So that raises the question, why are people exposing it to the web? Why do these people have such bad OPSEC?
Silas Cutler: I mean that’s a good question. So, when I first started trying to set up an instance that I could get screenshots and do actual testing before writing the blog, my default script set it up on a 2 Gb/2 CPU system and I couldn’t get it to finish the build process because of the system requirements. By default, the Docker instance will run the management interface on the loop back address so it wouldn’t be exposed, but I suspect it’s a lot of people that are deploying it on VPS servers, changing it so they can directly access it, instead of using something like an SSH tunnel to proxy their connection. In fairness, SSH flags are complicated. But we still see a lot. I mean I typically try to keep the kind of the gut estimate that for every one that we see that’s publicly accessible, there’s likely gonna be one that’s not publicly accessible. Maybe not necessarily deployed on something like a VPS, but even in even the lab in my basement, it wouldn’t be hard to deploy ARTEX and then use it against external targets. Of course it would look like my home IP address launching attacks, but there’s a lot of ways that it could be set up.
Himaja Motheram: Yeah. So what should people actually be looking for in their logs? Is there any way to actually track LLM assisted exploitation?
Silas Cutler: So there’s a couple ways that I’ve been hearing. The first thing that I’ve been doing, I used to write a lot of snort signatures, so I’ve been looking at a lot of the like at the traffic level, trying to pull out things like User-Agents. I haven’t seen it quite yet with something like ARTEX, but I’ve seen it with a couple of other self-developed tools by folks where it’ll actually put something like an email address in the User-Agent to identify themselves, but generally watching traffic, looking for user agents, sometimes it will catch things like ARTEX.
More often than not, it’s used like a Python User-Agent, but looking at net flow as well and especially for large traffic spikes has also been something that folks have using.
Himaja Motheram: Yeah, they typically do it on a pretty large scale.
Silas Cutler: Yep. They’re not quiet tools. I suspect if you’re using a fairly sophisticated model and giving good instructions to be cautious, it might reduce some of that traffic spike, but I think that all depends.
Himaja Motheram: Interesting. Yeah. I think my thinking around AI pentesting tools right now is that they’re still, you know, only as effective as the model that powers them and the human that powers them being able to find vulnerabilities that are again effective and can be exploited. Because I remember seeing some reports that ARTEX was being used to do a bunch of password spraying against just default logins and didn’t really go anywhere. Then they did that ConnectWise campaign and that was weaponized pretty quickly.
Silas Cutler: Yeah. So the first thing that I tested ARTEX on was a project that I run that’s written in PHP, without providing a lot of good context and just saying, we’re doing an audit of this site, it missed a number of things that I expected it would have found. It didn’t try searching pastebin and looking for a lot of the leaked keys that I know are out there. So there are limitations. Much of that seems to be both the model and the human creativity around it.
Himaja Motheram: Yeah. So you’ve been in the loop.
Silas Cutler: That’s right. I think we’re getting close to time. So I think let’s check to see if there’s any questions that have come through.
Silas Cutler: One question for today so far. Do you have any insight into whether the fridges are being hit directly, i.e., Modbus or some other IoT protocol directly on the Internet or via SAS portal from the manufacturer, i.e., weak credentials account takeover?
Himaja Motheram: I can take this one. So unfortunately all of the reporting around the exact initial access factor is still like seventy percent confident, maybe. But what we do know is that there were a couple of attacks that were not like an IoT protocol. This is the web accessible, like controller login interface to the fridge, to the physical device. In one case, I believe it was just like the default credentials had not been changed. So they were able to log in with an admin/admin type of situation. In the French blood bank situation, it was reportedly the orgs (likely) third party contractor that had some sort of vulnerability in their SAS portal. So a couple of different things, but not to my understanding, no real CVE, is mostly just hygiene stuff in the web accessible interface to the controller.
Silas Cutler: That makes sense. I know there’s been some discussions about attribution around these attacks as well being potentially Iranian origin. I know in the past we’ve seen threat groups from that region targeting things like MQTT. I remember a pretty well done GIF, that I still am not quite sure on the legitimacy of, but definitely looked like it was some MQTT setup that they had targeted.
Himaja Motheram: Yeah. I’m curious to see how that will unfold over the next couple of weeks.
Silas Cutler: Yep. One other question that came through, is there any way to track the refrigerator attacks and Censys? Yes. So we have fingerprints that we put out through the Danfoss blog as well. Using those searches you can look at what the current exposure is and walk back as well through those results to look at how those have changed over time as well.
Himaja Motheram: We’re also continuing to look for other kinds of OT refrigeration controllers to sort of get ahead of what might be out there that has not been touched yet.
Silas Cutler: I’ve been updating my eBay auto searches to try and look for some additional controllers as well. I was really surprised, the Danfoss controllers still look to be around like $600 USD. If you want to get a brand new one, it seems like it’s about $3200 USD. But more than what my budget is though. I’d love to just have this running in the back of my office.
Himaja Motheram: How cheap?
Himaja Motheram: Yeah, totally.
Silas Cutler: Well I think if that’s all the questions, then we can start wrapping up for the day.
Himaja Motheram: Great talking with you.
Silas Cutler: All right. Thanks folks. We’ll be back next month. Cheers.
Catch the Next Episode on Dec 9
Censys ARC Flash goes live once a month. Sign up to sit in on the next episode when it airs on December 9th to get the latest cybersecurity intelligence and ask your questions live.

