Latest / Along The Edge Podcast: Breaking, Defending, and Understanding Agentic AI / AI Security Brief - October 22, 2026
Transcript
- Andrius Useckas: Welcome to another news episode of Along the Edge Podcast. I'm Andrew Sushetskis. I've been in security for twenty five plus years and been doing AI security now, I've been told two and a half years, not two years. Just Alex: two and a half. Okay. Yeah. Andrius Useckas: Yeah, yeah. Time flies. Just Alex: Yeah. so I'm just Alex and this I've been I've been breaking and protecting AI for a few years now in in cyber for a while. Andrius Useckas: Some fun news this week. the big one obviously is the PHP core vulnerability we're gonna talk about. Then there was another news that you know, AI attacks are increasing against against different company websites. a lot of it seems to be coming from China and so on. And then we're gonna go back to our favorite subject, which which is Metas Muse and what it's doing now and security issues newly exposed. Maybe we'll touch on the certificate issue out of India as well, but not really AI related. So Just Alex: Not really a yeah, yeah. Andrius Useckas: No, it just stupidity. But okay, Just Alex: Pretty big though, but not a I. Andrius Useckas: we can touch on it quickly. So yes, when companies issue certificates to you, there are intermediaries basically that Confirm that you know certificates that you're using are valid or whatever. Just Alex: Yeah, it's authorized, yeah. Andrius Useckas: Authorized. And they kind of sit between the main security issuer and the customer. And they're all over the world. So this was one of the companies, e Mudra, I think, out of India, that basically issued private keys. All of these obviously are based on cryptography, on you know, put private public keys, all of that, you know, what's what's distributed usually with certificates or public keys. And private keys have to stay private. But in this case, it turns out that the way they generated those private keys was literally just taking number four. Number eight and a bunch of like basically up to two hundred and fifty-five, like one byte keys. and they got brute forced and obviously exposed a lot of certificates out there. As I said, not really AI-related, but I mean could probably be exploited today. It is brutal. Just Alex: Brutal though. Yeah, brutal. And embarrassing. Andrius Useckas: Yeah. Now PHP to shell. So that hit this week. Basically a vulnerability in the core of the PHP. That allows unauthenticated code execution, which basically exposes all the WordPress sites, all the Magento sites, and everything that runs under a PHP umbrella. Obviously, you know the news is saying this is AI again and everything else. Well, the truth is a bit more nuanced, I suppose. Yes, AI helped with discovering this vulnerability. But this is one of those bugs that have been in the Researchers' hands now for what, two to three years basically they've been working on it and longer, longer. Okay, so Just Alex: Longer, yeah. Two thousand since two thousand four would be the origin of this particular these two per two thousand Andrius Useckas: Two thousand four. Just Alex: four, yeah, was the original research. Andrius Useckas: I see. Just Alex: Yep. Andrius Useckas: Yeah, and i it it's just funny for me to see all the news outlets going, This is unprecedented, we've never seen this before and trying to put put it on AI. I guess they they forget or they don't care about things that happened, you know, not even five years ago, like, you know, Lock for Shell, that exposed millions of websites through Java library vulnerability that, you know, it resulted in what? Code execution as well. Just Alex: Yep. Yeah, Java Andrius Useckas: Yeah. So Just Alex: class was really easy to obfuscate. You could use LDAP, like all of these modifiers in the payload that were interchangeably to so like it just bypassed every firewall in existence. Andrius Useckas: Yeah, so it's pretty similar from that perspective, except that it's PHP. we have no love for PHP, at least I don't. I don't know about you. I I Just Alex: No Andrius Useckas: used to code PHP like twenty something years ago, and got rid of it as soon as I could. And platforms like WordPress scare me, so. Just Alex: Yeah, I would say I mostly avoid interpreted languages in general, exactly for these types of problems that can come up. Not just security related, but just bugs that you can get that are actually due to the interpreter that have nothing to do with what you're working on. and like the scope of this just to call it out. So we like you said WordPress, Drupal, Joomla, MediaWiki, RoundCube, Magento, PHP My Admin, Nextcloud, Larival, Simphony. I've never used that. WH C S, never heard of that. Don't know what those are. so like anything from five point four point zero, PHP five point four point zero on, which I believe came out in twenty twelve, which is wild. Andrius Useckas: And apparently they're not gonna issue patches for those, only to the latest version, eight whatever. Yeah. Just Alex: Yeah. Just forcing everybody to upgrade, which is just a huge nightmare for everybody. Andrius Useckas: It is a huge nightmare if you ever use PHP. Just Alex: Yep. Andrius Useckas: But but as you said, this is I mean, this is not really that new. not necessarily in PHP. I mean that there have been, you know, bugs in PHP itself in the core. But if you c if you take products, as you said, you know, interpreted languages, all of that, written on top of PHP, like WordPress. We've seen vulnerabilities in WordPress hit almost every week for the last fifteen years. So Is this unprecedented? I don't think so. Yes, it's a bad one. Just Alex: Yeah, it's terrible. And I mean, you called it out too, where it's like there was no I mean, the researchers themselves stated this. So this isn't just hearsay. this is directly from the researchers stating that there was no special harness used, there were no models used that were not generally available to the public. And they very specifically called out that the only reason they were able to come up with this even using an agentic harness, like an agent is not going to find this on its own. They very deeply understood. The history going all the way back to two thousand and four, with let me just find the name so I don't butcher it. Stefan Esser was the original Andrius Useckas: Okay. Just Alex: person who discovered that PHP memory limit remote vulnerabilities were an issue. And then I mean this had been being worked on by Caliph since August first of twenty twenty-six. August first. So I mean they've been they've been poking at this for a long time. And If they had not known that history and they if they did not know which very directed questions to ask the agent and kind of guide it into super specific areas, what they were stating is LLMs tend to start with like the most like the newest histories and like the newest piece of information and kind of work back from there. So even if you just let this thing run, it doesn't just discover these things on its own. But them knowing the history and kind of asking to go back and look at this and these initial findings that led up to this is how they found the initial bug and then like further guided it to find these two critical exploits. Andrius Useckas: Yeah, yeah, makes sense. What kinda makes me upset a little, I suppose, is that they slowed it down. So it was discovered in August by the looks of it. And Just Alex: Yeah. Andrius Useckas: who worked on that is the same cabal that has access to Mythos. So basically these select, you know, few large corporations like Cloudflare, like Palo Alto, that basically, you know, sat on this until they had signatures that they could deploy and use it for in the marketing campaigns. Which I know everything is fair in business, but come on man. It's like it's this is supposed to be an open community, all of that and then and you just gonna, you know, sit in on it until you get the signatures and then before anybody else gets a job on it, just use it in your marketing. I don't know. I just it just feels a bit dirty. Just Alex: Yeah, it is dirty. No question. Andrius Useckas: Anyway, moving on. Autonomous agents. Yeah. Just Alex: Patch now. Patch patch now. Yeah, sorry. Andrius Useckas: Batch now or take down your website, according to NSA. That was the advice from the NSA, which was kind of okay. Whatever, man. But that brings us to the next topic. So agentic attacks are increasing, and that's based on Nor Norge's research and Reuters basically published an article based on that saying that they've seen a ton of agents hitting different, you know, what corporate websites and government websites. Using agentic flow. So just basically somebody told an agent what to do. And it just goes and basically hammers the specific resource with all kinds of vulnerabilities until it finds something. Again, again, not a new technique. I mean, we've seen botanists do this all it's just AI driven at this point. And yes, it's more dangerous. You need to detect it. It's a different kind of, you know, beast. It moves faster, but I mean, we've seen this before. It's just faster. But aga I guess my point is that with AI doing helping with the research and these vulnerabilities and then AI running the actual agentic attacks, we're probably gonna see more of these kind of things like lock for shell. And we do need protections that can detect it instead of this cabal of you know a few organizations that with a special subscription that you can subscribe to these, you know, signatures, virtual patching signatures that will be released just specially for you before the actual you know exploit is released. We need an actual solution that can find this real time, or at least within seconds and not within two months. Just Alex: Agreed. A very challenging problem to solve. Andrius Useckas: It is a challenging problem to solve. But if AI is able to attack, AI is able to defend. I mean, if you look at lock for shell, it's not just some obscure vulnerability that you have to, you know, do incantations on it to get the shell and everything else. You're still using PHP code in it. You can still detect the actual behavior. And anything, you know, AI driven should be able to do it. I I know, you know, you cannot use AI line speed, but I mean there are solutions how to pass it through layers of heuristics and things like before you get man. Anyway, moving on again. So Muse AI. again in the news. Apparently they had some feature where they allowed you to share the entire conversation that you had with the DM. And then people started posting Just Alex: Yeah, and this was the Andrius Useckas: it. Just Alex: the second time that I'm aware of that a sharing of AI conversations led to a massive leak of information. The first one being open AI. Andrius Useckas: What was the exact information that was leaked, do you know? Just Alex: For open AI or for meta? Andrius Useckas: No, we can talk about Meta. I mean Open AI we can touch on. Just Alex: meta, no, I don't know. Andrius Useckas: Yeah, from what I understand it was yeah, like company, like privileged information basically that people put in Muse AI and then basically share those conversations with other people within the company and thus exposed it. Just Alex: So it was like cross conversation leaking or was it like once you share Andrius Useckas: No, so it's just Just Alex: it, it was publicly available on accident? Like that that was exactly like what happened with OpenAI. Andrius Useckas: That's exactly it. That's exactly it. So you like you talk to your muse AI and if you're in marketing or something like that, give it like privileged information about the company's finances and things like that. And then you decide to share the entire conversation. And because it is available on the web, it's exposed. Just Alex: Hmm. Pretty straightforward, I guess. Andrius Useckas: Pretty straightforward, I guess if you want to share private information. Just Alex: You just need the I mean you it's not like a numer you can't enumerate for this information, I assume. Like it's not like a walkable key. I Andrius Useckas: I assume not. Yeah. Just Alex: mean, but it's so somehow you have to Andrius Useckas: I don't have to yeah. Just Alex: get you somehow you have to find that endpoint one way or another. Andrius Useckas: One way or the other, absolutely, but it was a big enough issue for the meta to disable that functionality. So that's all I know. I did not dive too much into it, obviously PHP for Shell took most of this took most of my attention this week. Just Alex: Yeah, same here. I was reading the whole history. It's actually really interesting. Andrius Useckas: Yeah. It is pretty interesting. All of these like you know, dots and Muse AI, yes, we already know that there are security issues there. So same advice. Just Alex: Mm-hmm. Andrius Useckas: Don't put your real information in there. Private information especially. And don't give it access to real email addresses. Just Alex: Yeah, or your bank account, your bank cards. don't don't share your medical history. Don't let it plan your doctor's appointments. Andrius Useckas: Absolutely. Just Alex: Stuff that just feels really obvious, but I'm sure people are doing it. Buying buying groceries and shopping Andrius Useckas: I'm sure people are doing, but yeah. Just Alex: for the best deal on something. Andrius Useckas: In conclusion, this is a new world. AI is definitely gonna be prominent in it, I suppose. From discovering these vulnerabilities as well as attacking corporate resources and so on and so forth. It's not gonna get better, it's gonna get worse. We know that much. It does not but at the same time, you know, do not buy into the sensational stories from the media. we in cy cybersecurity have been dealing with these issues for a long time. It's just kind of nice to see them finally bubble up and people actually paying attention to it for a change. My only hope is that it's gonna translate into, you know, boards actually approving finances for security teams and not just them running on bare bones, as has been the story the story for the last twenty years. Anyway, good good segment. we'll see you next week. Thanks for listening.