Latest / Along The Edge Podcast: Breaking, Defending, and Understanding Agentic AI / AI News Brief - September 17, 2026
Transcript
- Andrius Useckas: Welcome to another news segment episode from the Long The Edge Podcast. I'm Andrew Sushetskis. I've been breaking AI agents now for two years. Alex: Hey, I'm Alex Gatz, staff security architect, and I've also been breaking AI agents for more than two years. Andrius Useckas: Yeah, so there's been a lot of news lately about, you know, how AI is gonna kill us. Speaking of, you know, agents and breaking them and everything else. Some say by the end of the decade, some say in ten years. And what bothers me isn't the claim. It's what the claim does to Durham. Everybody picks a side at this point. It's so polarizing for some reason. It's like people don't want to have a discussion anymore. It's either the camp of, you know, yes. It's gonna kill us. And then, you know, they give us a bunch of these science fiction, almost like science fiction episodes, as to how it's gonna happen, how AI is gonna grow legs and how it's gonna duplicate itself and find itself GPUs and everything else. And then then there is the other camp that says everything is fine. We have the controls to do it, and nothing is gonna happen. And here is this thing that I actually keep noticing about who is talking. It's it's not the you know cybersecurity professionals. It's not the guy that worked in the actual BSL3 lab before and knows exactly how to engineer viruses and stuff. It's pre-training researchers, lawyers, politicians, and the occasional podcaster. So it's not what we're seeing from the actual professionals out there. And let me be clear about that. You know, it's I'm not saying that AI is not dangerous and people are stupid. What I'm saying is that. Yes, AI can absolutely be dangerous in the right kind of hands. It what it is in the end is just it is an instrument. And who uses that instrument matters. It is just like a hammer, you know. It depends who uses it. I mean, we've been using AI agents and AI models now for two years to do all kinds of offensive work. That that is what we do. we pen test agentic AI with our own AI. We've seen what it can do. And and some of that stuff is pretty cool. I mean, if if if you use like an uncensored model that actually lets you do all of these things, actual attacks, you 10x, you 100x of pen testing work. It it's great from you know pen tester perspective to find these different gaps and so on and so forth. But the model also hallucinates. And from the attack perspective, that is not necessarily a bad thing. I mean, hallucinations allow us to find. Specific issues that a human would never find. So can agents be dangerous? Absolutely. but alignment is not a way to control them. You cannot just nicely ask the agent, please do not do this and do not do that. I mean, even anthropic themselves, they have all of these set cybersecurity controls in place. Actual definitive controls. What is that framework that they use? ASL3? There is actual controls in place that are outside of the model. It's not an alignment problem. Yes, alignment is important. We agree with all of that. But hey, you gotta listen to the actual cybersecurity people just like Entropic does. And as far as cybersecurity people go, you know, stop releasing these reports that look like marketing stuff. There is no actual substance to it. I mean, Alex, we know what RCA is supposed to look like. What does RCA look like? I mean, timelines. Alex: Yeah, very, very specific payloads that were sent, very specific notes on what infrastructure was involved, very specific. We're missing all none of these details are here whatsoever. None of the details about like what was the Andrius Useckas: Absolutely. Alex: the framework that was used, what like what was the configuration of the sandbox that was involved, what was the the goal of the the agent and like what was it doing and what steps did it take? Where like show me the logs and the timeline. Andrius Useckas: Yeah, absolutely. And and just you know, releasing these reports, even from Metter w whatever you pronounce it l as, you know, those reports are interesting and everything else, but they are not cybersecurity reports and they're taken as one as those. I mean from Matter side themselves, apparently they just hired a cybersecurity lead on August thirty first. And the actual job posting says that you're gonna be the first cybersecurity guy in the organization. How how Alex: I didn't know that. Andrius Useckas: Yeah, how c how ca how can you be in this game, how can you be a a third party independent evaluator if you don't even have cybersecurity people on your staff? Alex: That's crazy. I didn't know that. That's actually really wild. Andrius Useckas: Yeah, I mean obviously I'm a bit triggered by all the you know posts on accident last week and everything else, but we we're not w we're trying to pick the middle path. We we are the cybersecurity engineers, we see what the models can and cannot do. We do not assign science fiction properties to them, we see just you know where the limits are. But Alex: Yes. Andrius Useckas: we also know what you can do with them. And And you have to have actual physical boundaries. It's not that complicated. I mean, if if your model can really, you know, bypass everything, like they say, which I don't buy. I mean, like, look at the WordPress attacks. Millions and millions of sites out there run WordPress. And there is at least like one vulnerability discovered within the plugins every week. But yet we do not yeah. Yeah, but but Alex: At le yeah, I yeah. At least Andrius Useckas: it's still old school botanists trying to exploit them. There is no Coordinated AI agentic attacks. If AI wanted to attack the internet, WordPress would be where it would begin. And we're seeing none of that. What are the lock for shell kind of, you know, attacks? Why are not we discovering more of them? That's I mean, doesn't that speak to the limitation of these models at this point? Alex: To your point about hallucination, it will make a lot of noise and like try to brute force its way in if you like if you just pointed an agent at it, and I mean you can put restrictions around that, but these would be like in code restrictions around like request rates being sent to the thing. no, I don't I've not seen an agent on its own come up with a novel exploit. I can say that for Andrius Useckas: Absolutely. Alex: certain. I'm n not like a completely novel exploit that wasn't a duplication from some other previous work somewhere where it's just brute forcing the same types of techniques. Andrius Useckas: Yeah. I guess my ask in the end is if if these labs are seeing something that we're not seeing, that you know, we've been doing this from cybersecurity perspective for two years, please give us actual evidence and not just market where. Give us the actual RCA reports with all the technical data, with all the timestamp technical data that is involved in it. Not just give us, you know, user as this and and then you know the model asterisk hallucinated back and prompt injection. Alex: Yeah. It Andrius Useckas: What does that mean? What is the context? What was the context? What were the prompts in between? Nothing. It's just like Alex: Yeah, it's literally just a wall of markdown, which is Andrius Useckas: Yep. Alex: just makes it look like the reports are just AI generated garbage, really. There's not any substance to any of them. There's no nothing I have seen looks anything like any other previous C V disclosure or like any kind of security report, any kind of RCA. even like a like a technical deep dive into an existing problem from a third party. I've not even seen anything interesting like that showing the Andrius Useckas: Yeah, it instead of all of that technical stuff, what we get is science fiction. Like AI is gonna, you know, just replicate itself. Adnauseum. Nobody understands what an actual, you know, frontier model, what what is the size of the actual frontier model? And even if you Alex: Yeah. Andrius Useckas: can copy those, you know, terabytes of data in that case, what are you gonna run it? How is it gonna get those GPUs? How is it gonna gonna get compute? I mean, people just don't think about the those things, and then you just get like On the other side, you get just funny comments like well, the labs will now have to like isolate, detests, and run synthetic internets. I think Andrew Young said that. What is a synthetic synthetic internet? I mean, it's it's just air-gapped network segment. We've been doing that for decades at this point. Alex: Yeah. I don't I don't I to me, okay, at a at a maybe too superficial of a level, maybe not, but a lot of the stuff that I saw on X and a lot of what you're talking about is regurgitation from like Nick Bostrom's book, superintelligence. I don't remember that came out a very long time ago now, like in twenty fourteen maybe, something like that. Andrius Useckas: Yeah, I think Elon was contributor to that as well. Elon Musk. Alex: So then a follow-up to that was Max Tegmark's book, Life 3.0. And both of these books, like Max Tegmark quoted Nick Bostrom's book many times in there, where they're worried about a slow takeoff, a fast takeoff. Like it was a bunch of pontifications and theories which are truly are just science fiction of like the different ways that AI could run wild. And all of what we're talking about here and all of what I'm seeing on X was already covered in these books. So to me, it just feels like We're just restating the same things that someone else already imagined, twelve years ago now. like Andrius Useckas: Yeah. Yeah, and look look w yeah, I I'm not saying it's not possible. I mean, science fiction things do come through. We have, you know, robots coming up and things like that. But at the same time, Alex: Sure. Yeah, certainly. Andrius Useckas: if you start with the reasonable controls and don't just rely on the and l on alignment and actually involve cybersecurity engineers or you know, people that build the labs for the biologists for the biology stuff or of that, which I know nothing about, so I'm not gonna talk about it too much. But you know, there are actual experts that've been doing this for two decades, three decades sometimes, you know. Why don't you consult them? Wha why are you just pushing this stuff that, you know, you're trying to dress up as cybersecurity when all it is, all it looks like is marketing? Alex: you could argue if these agents are so successful, well, what is the equivalent to that that we've had for the past decades would be like nation state level advanced persistent threat type stuff, right? So, like we've had strategies for those cases for decades, as you say, and like those Andrius Useckas: Absolutely. Alex: cases obviously have to evolve. And that is I mean, it really is an un that to me that's kind of an unsolvable problem almost, like on that side. But Let's say like that's the top tier, like that's as good as it can possibly get. And we already have strat strategies in place for that. So are we saying AI is going to magically become better than the best of the best in the world that has unlimited amounts of funding and resources to back them up? Is that is that the argument we're trying to make? Andrius Useckas: I guess so. I don't know. I mean some people compare them to nuclear weapons, basically. But even for nuclear weapons, we have the protocols. We know how to restrict all of that stuff. and yeah, science fiction things like, you know, the model is gonna come up with a novel virus. It might come up with, you know, a thousand different viruses, but you still need the actual lab. The an actual lab to actually implement it and test it and everything else. How many people in the world can do that? And then you have to source all the materials and you basically have to have you know millions or billions of dollars in addition to the actual theoretical stuff. I mean, on the other side, people say the it's also gonna cure cancer. Yes, it can you know come up with the new proteins and new ways to battle things as well, but you still have to go through the same process that will take five, ten years or whatever else. You cannot just take all the process out just because hey came up with something. It's an interesting world, an interesting discussion. But I think, you know, in the end my point is that it needs to be a discussion. Don't just pick a side and and try to die on that hill, so to say. It's I don't know, people just don't like discussions anymore, I guess. I do. I still Alex: No, it it is you keep seeing this argument where it's like if we over regulate, we're gonna fall behind other countries. If we under regulate, we're gonna have a misalignment problem. it's it's very hyperbolic. Like people are very strongly opinionated in one direction or the other and they're Andrius Useckas: Yeah, but but you you can do things. Again, all of the there are solutions to all of these things. Yes. So you you say okay, you cannot release this model to the public because it's too dangerous. But you cannot fall behind China. So yes, you can set up the entire air gap data centers in that case. Where you can run all of these experiments, you can have all of these very dangerous models. They cannot escape. There is no Wi Fi. I mean you you can, you know, paint the entire building in copper in in copper paint and then you know you Alex: A fa a Faraday cage? Make a Faraday cage? Andrius Useckas: Yeah, Faraday caged the entire data center. I mean, it's been done. I mean certain, you know, DOD contractors have done it to the entire buildings. It's not that Alex: yeah. Andrius Useckas: yeah, it's all of that can be isolated. If you s if you say it's so dangerous that you cannot continue the research, that's BS. You can still continue the research. You just don't need to release it. You can, you know, convert it into just like you know what military does with their research. To different levels of sensitivity, to different levels of controls. Alex: Yeah, I mean what are we saying it's so dangerous that even if we do that, that the AI is going to blackmail or extort someone to convince them that they have to like take it out of the building and run like provision provision new resources. Andrius Useckas: And how they're gonna do it? I mean they they can't put it on the phone. Well, I mean you can also take the phone. Yeah, exactly. How they gonna do it? I mean, yes, you can Alex: No. I don't know. Andrius Useckas: get into science fiction scenarios at that point, like you know, lasers or over long distances, it's gonna like you know, create its own robot lab or something like that. But but that's you know, that's science fiction at all. Alex: I think we would I would hope I'd like to think we would notice if that much money's being moved around. Like make fake this Andrius Useckas: No kidding. And at that point we Alex: is now we're theorizing, but like make up fake people, make up fake people and then Andrius Useckas: Yes, we are. Well but it's kinda fun. It's a fun discussion. Yeah. Yeah. Alex: like make a fake company out of fake people and Andrius Useckas: But it has no internet connection, it cannot do any of that in that case if it is isolated. And in the end, if it even if it finds a way to somehow do something, we will notice. And we can just cut the electricity. It it's not that complicated. So Alex: just persist across the entire internet, like I don't Andrius Useckas: Yeah. Alex: know, getting into like the blockchain space, like running it ex it it exists through blockchain. I don't know, like use it let's use all the the keywords. We got blockchain, we got quantum, quantum cryptographic functions being used Andrius Useckas: There you go, it's gonna Alex: and Andrius Useckas: it's gonna use quantum entanglement to entangle itself to another internet node and escape it that way. Alex: Ha ha ha. Andrius Useckas: Anyway, okay, enough pointation at this point. Our point is discussion needs to continue. And there is a middle path to this. There is no like, you know, one side or the other. if we just split and polarize ourselves, we're not gonna win this. We need to figure out actually how to apply proper controls, and in that case, cybersecurity and you know, biology engineers are still you're still b best better. All right, thank you.