NVIDIA Solves Rogue AI Agent Problem - OpenAI is DEAD

Eli the Computer Guy (0:00)

Coming from CNBC, Nvidia releases a software platform to stop AI agents from misbehaving.

And I think the words I could use for this is it's about gosh darn time. One of the things that I found fascinating with this whole AI revolution is basically how little security has been involved. And it's weird because this revolution has been going on in the 2020s, and we've understood security for a long time, right? And so it's weird that so much of this stuff is being deployed with a mentality of security that really reminds me of circa 2001, not something we should be looking at at 2021 I've talked about this before where I was at a meetup a while ago, maybe it was about a year ago when agents were really becoming popular, and I was with OpenAI and Research Triangle, or not, OpenAI, oh, All Things Open. It's called All Things Open and Research Triangle. It's a phenomenal meetup group. Anyways, I went to a meetup, and there was this guy and he was showing this whole coding system, basically agentic coding or whatever else, and he was showing how the system was able to interact with this shell. So interact with the shell, so that means it could create files, it could create folders, it could delete files, it could delete folders, it could create databases, and my friends, it can delete databases. So anyway, as I was watching all this, technically very interesting, so at the end, I raise my hand, like, hey, yeah, so it's an old school MCSE, I got a question here.

So normally when we think about security on systems, we think about the user account. You log in with a user account and you get a certain level of what are called privileges, that is what you're able to do on the system, or what you're able to do on the network. One of the problems with this particular way of doing things then, is anything that is fired off, any action that's fired off during that session, will then be able to be fired off at the privilege level you have. So if you are a root administrator on a system, it means anything that's fired off, whether it's you, whether it's malware, whether it's an AI agent, is able to do anything that root is able to do. So this is one of the things you go back to security, and old school security from 2004, and all of your users shouldn't even be able to add printers to their computer. If your users can't install anything, they can't access the registry, they can't install a printer, they can't do anything, then even when you fire off a malware attack, a spear phishing attack, the spear phishing attack is gonna fail because you don't have permissions to do whatever that script wants to do. Again, this is where we're talking about the layers of security. That security should be an onion if any layer of the security is compromised. The security still should remain attacked because there are multiple layers there. Somebody clicks on the spear phishing email, that's a vulnerability. Somebody clicks on the link in the spear phishing email, that is a chink in the onion. But when the script fires off, it fails out because the user account doesn't have permission. One of the interesting things, though, is when you have coders, coders are doing a lot on their system, and so they want to have root access. And so one of the things I was asking is, so what do you got? What is the Active Directory? So that's when we talk about privileges in the Microsoft world, something called Active Directory. Woo woo, I'm a fan of it. Anyways, what is the Active Directory of 2025? What is the Active Directory for this AI agentic world? Again, to kind of separate out the privileges of the user versus the AI system that the user is interacting with. Again, something to be concerned about. And yeah, it was not a good answer. Oh my god, it was not a good answer. Anyway, there's like 100 people in the audience. This is Research Triangle, right? Real tech scene. You know, this is a long standing event, whatever. And literally, the response from the guy is like, well, you trust your developers, don't you?

And then I shut up. Then I was like, thank you for your answer. Cause if I had said any other words other than thank you for your answer, it would have been about five minutes of me screaming swear words at the top of my fucking lungs. Cause like, what do you mean do I trust my, I know, I don't fucking trust anybody. I don't trust my own mom, and she's five feet under the fucking ground, right? Again, as I've taught about before, distrust, verify, stay suspicious. Do not trust anybody for any, like you just don't trust. You build the systems, trust the system. Don't trust the fucking people, right? Anyways, and I found that weird, cause again, that's not from even 2020 or whatever else. That was sub a year ago. A real professional developer was talking about that. And I think that's one of the kind of interesting things when you start hearing about these AI agents. One of the most interesting things is just basic, basic system security, right? Like we hear about this with these vibe coders crying. Again, we've heard CEOs, CEOs of tech companies crying when the vibe coding platform wipes out the production database. And we have to understand about why a CEO crying about the vibe coding system wiping out the production database, why you have to understand that the CEO should not just be fired from their job, but all equity they have, even if they're a founder. Oh, what? You spent the last five years working 24 hours a day and 30 day months to get this company where it's at? Well, then I guess you have the requisite skills to start the fuck over again from square one. A, you're going to be fired. B, we're just going to take that equity for your stupidity. But the thing is, like when they talk about wiping out the production database, right? File permissions should not allow that to be able to occur. But it's weird. It's weird, right? These vibe coding platforms wipe out the production database. And not only does that happen, and like the weirdest part about it is, is that the CEOs or senior developers or whatever else, then go on to Twitter, then go on to Ars Technica, and with big tears, they start crying because of how bad these vibe coding platforms are. And I'm like, what?