Latest / Along The Edge Podcast: Breaking, Defending, and Understanding Agentic AI / AI Security Brief - October 2, 2026
Transcript
- Andrius Useckas: Welcome to another news segment of Along the Edge Podcast. another busy week in AI security and news and all that stuff. I mean, stuff coming out of the White House and Anthropic releasing an article on GLM 5.3 and how it's a dangerous model. and another assistant released this time by OpenAI. Anyway, let's talk about it. So let's start with the Exciting renaming of the AI into SI. What is it? Artificial intelligence to superintelligence? W w what is that even supposed to mean? What what what w how does that change anything? ex Alex: I have no idea. It someone's gotta someone's gotta change the labeling on a lot of documents in the in the in the government. They're gonna be going through whiting out lines of paper, rewriting super intelligence over artificial intelligence. It's gonna be a whole process. Andrius Useckas: Yeah, but is it even gonna change the term? I mean, people gonna use SI at this point. I'm not gonna use SI. Alex: No. No. Andrius Useckas: SI to me is again like Peter F. Hamilton books and his what what was that trilogy? Salvation trilogy, that they had a side there which was sentient intelligence. So for me, to I will always associate it with that specific science fiction trilogy and sentient intelligence, not some super intelligence. Alex: Is it super intelligence? Is that what they're renaming it to? Andrius Useckas: Yeah, yeah, superintelligence, that's what it is. Alex: That's so weird. Andrius Useckas: But if you own a Slovenian domain with dot SI at the end, basically, I guess you can make some money now. Alex: That's tr that's actually true. Yeah. I mean the the the the rates are going up. They're gonna go up now. Andrius Useckas: Speaking of government doing things, I guess there be like w what what is the agency agency investigating and dropping an open AI at this point? The FTC. Alex: The FTC, which is weird to me. Yeah, yeah. Andrius Useckas: What what is the story behind that? Alex: the entire focus was like concerns around danger. I'm quoting an article here. This is AP News. the the quote is concerns around the dangers powerful AI models may pose to humanity have led some of the top AI developers to call for a slowdown. So they're apparently they're doing it because Dario Amade and Sam Altman were calling for slowdowns, I guess. Which Andrius Useckas: I guess. Alex: just like to me, that just solidifies their position in the market if everybody else has to slow down. Andrius Useckas: Yeah, well everybody else is not slowing down. And speaking of which, GLM five point three model has been out for a while now, I guess, a couple of weeks, but Antropic released a Alex: Yeah. Pretty good. Andrius Useckas: an article about it saying it's a powerful model. It does it can find or basically create all of these vulner exploits for vulner vulnerabilities. They compared it to the Mythos preview, I guess. And GLM was able to create fifty out of four hundred and ten attempts versus Mythos. That was able to do fifty six out of one hundred ten. Alex: Yeah, it's it's pretty good for security use cases. It really means. Yeah. Andrius Useckas: It it's fantastic. I mean we use it. It's great for defense. And that's why I mean w that was the entire thing in Open AI and Hugging Face story. Hugging Face used GLM because they could not use Claude. They could not use OpenAI models because of all the guardrails and restrictions. GLM does not have as many, so you can actually use it as an effective model. So I I don't know what the purpose of this article was. It it looks like advertising to me. It is like cannot use us. Cannot get access to mythos, use GLM. Alex: I think it's slander if anything. I think they're trying to basically say like, Hey, you shouldn't use this. maybe they're trying to get it blocked in some capacity or they're trying to scare people off from using it and saying they should use Mythos instead. Andrius Useckas: But they're making kind of the case weaker because they're saying this, you know, extremely powerful model has been out now for a while and the world has not ended. We have not seen an increase in attacks. It's Alex: That's true. Andrius Useckas: we have not seen an increase in vulnerabilities and so on and so forth. So I don't know what the case is. I mean, yes, it's a great model to use in defense because I cannot we as a startup, security startup, cannot get get access to mythos. The only access to myth you have to be a large company. Alex: No. Daybreak even. We can't even we're we're having trouble getting into daybreak. Andrius Useckas: yeah, it's been three weeks like attempting to get daybreak. I finally got a hold of someone on X that works at OpenAI and they're trying to fix it. It's been what five days now? It's not fixed yet. It's like even internal guys cannot fix debugs. It it all Alex: Yeah. It's a joke. Yep. No, it's a total joke, honestly. Andrius Useckas: I know. I mean they require you to buy Yubi keys now to have access to daybreak. Do you buy Yubi keys and then everything breaks on the actual verification? They use Persona. I think that's what the solution is called. And it's it's an absolute disaster. It it looks like a vibe coded platform that nobody really like, you know, tested or thought about and I would be embarrassed either. They don't seem to care, but I would Alex: Yeah, was the was the note true? I this was like hearsay, so like allegedly open AI has some kind of stake in Yubico. Was that true? Did you ever look into that? Andrius Useckas: I did not, but yes, they send you to you the webpage. You have to buy Yubi keys. Luckily I have a ton of YubiKeys, so I did not need to buy any to actually Alex: Same. Andrius Useckas: get stuck on the step number three. But anyway, you can tell you can Alex: Yep. Fight two base, I believe. Andrius Useckas: tell them a bit about that. Let's move on. NVIDIA I guess releases the security platform for agents. What's that all about? Alex: I actually like this approach and we like we have talked about this in the past about moving the protection to like the hardware layer so it's not just all in code. so I don't know, they kind of called it a quote unquote kill layer for AI agents. So they it's called open shell. It's an open source runtime that enforces file system network credential and tool access policies outside of the agent itself. Yeah. Andrius Useckas: So so there are two different parts to it, right? There are two different parts. There is Open Shell and there is the actual hardware that does the monitoring and enforces, as you say, the kill switch hardware. Alex: Yeah, blue field four hardware. Andrius Useckas: Right, right. Alex: Yeah. Andrius Useckas: It's an interesting idea, but who's gonna pay thousands and thousands of dollars to run the skill switch monitoring? Alex: Probably the government or other air gap systems that are or like more critical systems that wanna use AI but are scared too right now. I would I would assume. Andrius Useckas: Yeah, but you could do Alex: That would be the that would be the use case, right? Andrius Useckas: It would be, but you can do the same monitoring with software, so I'm not really sure why you need this in a hardware solution. Alex: 'Cause I think the fear is the sandbox escapes and the all all of the I d I wouldn't call it hype. It's like all the negative press around how easy these agents have been finding it to get out of their test environments. However, like we know by looking into these that the reality is we're re the reality is like these were misconfigurations or they were specifically set up to be have direct access to the internet in the first place. So like the restrictions weren't even there to begin with, but it was being like hyped in this other direction. So fear mongering, fear mongering really. Andrius Useckas: Yeah, speaking of sandbox escapes, I read some Yeah, it is. But I read something funny on X today about the escape, like some organization, Pawn Box, something like that. I don't know. But Pawn Machine, maybe? I don't know what the name was. But they yeah, they they basically deployed like Alex: That sounds familiar. Pwn box sounds familiar. Pwn box. Andrius Useckas: four agents and containers and gave them instructions to escape the sandbox. So basically like in a containerized sandbox, whatever. Then apparently they tried Thousands of exploits, they wrote a bunch of C code, did not escape. All they did is kill the host through all OM attacks, out of memory, but not attacks, out of memory issues. So I was Yeah. Yeah, exactly. Alex: kill. So they just killed it. They made it kill. Andrius Useckas: The Yoom killed it. But but the issue is if you're running a thing like this, don't you know that you can assign like resource restrictions to these containers so it does not kill your server? Alex: Yeah, no kidding. Yeah. Yeah. I mean that's like fundamental container runtime usage in general. Yeah. Andrius Useckas: Indeed. Alex: Docker you have flags for it, Docker compose you can define it right in the YAML file. It's all it's all over the place. Andrius Useckas: Absolutely. I I mean we know it. Anybody with any kind of DevOps knowledge knows it. So it's like saying, Alex: Yes. Andrius Useckas: this might it might be and it's on attack if you actually kill the host through the you know. It's okay, no, it is not. You just basically assign resources, constrained resources on the container and you're fine. Speaking of containers. Absolutely. Alex: That again falls into misconfigurations for me. Like that's a misconfiguration. Or, you know, like lack of knowledge around how to even properly do DevOps. Andrius Useckas: Absolutely. Speaking of all of those VMs, misconfigurations and so on and so forth, Muse AI is in the news again. reading private messages, Alex: Yeah, yeah they are. Andrius Useckas: what was the other story? I don't recall but Alex: man, this one was funny. So there's a YouTuber, I'll call him out. So Matt Rob. He was using Muse to accept a marketplace offer and it supplied the buyer with his home address, and the buyer arranged like showed up to his house. The buyer actually showed up to his house. And he like Matt Rob had no idea about this whatsoever. No, it's not really it's a little I mean, okay, but Andrius Useckas: I should not laugh at this, but yeah. No. It Alex: like don't give it don't give Muse your address, first of all. Andrius Useckas: Agreed. Alex: Or like access to any kind of personal information. But like we know that, of course. Like we understand the the implications of these things. but if you're just using it how Meta advertises it, yeah, it probably is gonna get your address from your like email profile, right? If you filled out your whole email profile, which I personally never do, but a lot some people do that, right? So yeah, it just it just showed up right to us a random stranger showed up to his house knocking on the door. Andrius Useckas: Yeah, not as interesting story, but also related to Muse AI was that some guy, I don't remember his name now, he published an article basically. He asked Muse to publish to to create some kind of story or some kind of article. And he noticed that a lot of things or some things that came in in that article were related to his private messages that Muse was not supposed to get have access to. He never gave it permission to do it. So you did some tracing and yeah, you figured out that Muse actually read his private messages that never gave permissions to do any of that. And that is the issue with these personal assistants. I don't know. I mean, you can't really trust what you're doing yet. Alex: Yeah, I mean that that falls into an issue with what Meta is advertising as their extra security layer for these agents, right? I that Andrius Useckas: Yeah. Alex: that should have been picked up by their privacy, I don't remember what it's called now, but like that that privacy layer that they added that's supposed to be judging information to determine whether or not it's sensitive, personal, PII or otherwise. Andrius Useckas: Yeah, absolutely. And we know there's gonna be false positives, and that's what we're seeing here. Alex: of course, yeah. Andrius Useckas: That's why you should not give these agents your personal data and that's not only Muse AI at this point. I mean OpenAI at release dots now. very inventive name, obviously, but Alex: I think it's sh is it just like open claw rebranded basically with some some extra work on top, right? I mean they hired they hired the open claw guy a while back now. So I would think that Andrius Useckas: They did, they did. Alex: I would think that this would be this would be the thing then that comes out of that. Product the productized version Andrius Useckas: Well supposed to be secure. It Alex: of that. Andrius Useckas: is, but it's I think it like Muse AI as well, it it runs mostly in the cloud, so VMs and things like that. But but same issues. So all the tool integrations that it it it it expects you to do in order for for it to be useful, like you know, email and and all the other stuff. So yes, same kind of, you know, concerns. Alex: I think that I think it's bigger. It was bigger for me when I was reviewing this because it integrates with everything. And I mean it like pushes you to integrate with everything. And my my red flags are popping off around indirect prompt injections all over the place. I'm just you have like the like the very in-depth Slack integration across all of your chat messages. You got the the email integrations, voice base. I mean, there's just so many layers there that are concerning to me. I Andrius Useckas: As we keep saying, yes, this is cool new tech and it's going in the right direction, but security is still being overlooked. So if you do want to play with it, do so, but do not give it your main accounts. Give it like, you know, throw away email accounts just to test with. yes, it sounds nice to have an assistant that to do stuff for you, but you probably don't want your data leaking or somebody showing up at your door. Alex: I don't know if it's X or Elon Musk released that jokesite like dot it's like D O T and then a period com. So it's like dot com domain. Andrius Useckas: Dot com. Yeah, yeah, yeah. Yeah. Alex: Is that real or the is that a real agent or was it actually just Elon trolling Sam Altman and Andrius Useckas: No, it's Elon trolling OpenAI basically. So when they re when they announced, or I guess you found out before the announcement, you know, that they're gonna name the assistants dots, you basically went out and spent, the rumors are twenty million dollars to buy that domain, and then just forwarded it to Grogbot basically. So when you go to dot.com, you're just you end up at Grogbot page. So yes, absolutely trolling, you know, yeah, whatever. Alex: That's that's just so epic. I mean like Andrius Useckas: Yeah, I wouldn't call that security news, but it's kind of funny, yes. Alex: Yeah, no, that's definitely not security news at all. Andrius Useckas: Anyway, do not give agents all of your data. That's the final message for this episode. Thanks for listening. See you next week. Alex: Yeah.