AiCyber.Land

Shelby discusses Nvidia’s reported Hugging Face acquisition and the SafeMind red-team and defensive models. Bryce discusses GPT-6 ‘Astra’ and its cybersecurity benchmark results, followed by a conversation about misconfigurations, remediation, and testing AI’s ability to secure servers. They wrap up with MIT OpenCourseWare and a LEGO Charmander build.

Show Notes

Sources & transcript: Episode #43 page
Read transcript: https://share.transistor.fm/s/824d8f16/transcript
Apple Podcasts web strips note links: use its existing Episode Webpage action for clickable sources and the transcript.
Automated transcript; may contain errors.

Shelby discusses Nvidia’s reported Hugging Face acquisition and the SafeMind red-team and defensive models. Bryce discusses GPT-6 ‘Astra’ and its cybersecurity benchmark results, followed by a conversation about misconfigurations, remediation, and testing AI’s ability to secure servers. They wrap up with MIT OpenCourseWare and a LEGO Charmander build.

KEY MOMENTS:
00:00 - Opening & Introductions
01:02 - Nvidia's Reported $12.9B Hugging Face Deal
02:57 - Meet Red Tempest & Blue Solano: AI for Hacking & Defense
06:39 - GPT-6 "Astra" Released: Crushing Security Benchmarks
13:23 - Why Misconfigurations Are a Bigger Threat Than Exploits
16:15 - "It's Not a Vulnerability, It's a Feature!"
17:38 - How to Get a Free MIT Education (Seriously)
20:30 - Unwinding with a Smart Lego Pokemon

Subscribe and comment if you’d like to see server-security tests.
https://aicyber.land

Sources/context added Nov 9, 2026; selected-topic references, not verification of every claim or identification of the original on-air references.

Sources

  • NVIDIA: Agreement to acquire Hugging Face (2026-09-03). NVIDIA's announcement of an agreement to acquire, not proof the acquisition had closed or corroboration of the separate GPT-6 benchmark discussion.

What is AiCyber.Land?

Join industry experts and thought leaders as we dive deep into how artificial intelligence is transforming cybersecurity, shaping defense strategies, and creating new opportunities in the digital landscape.

Automated transcript; may contain errors. Formatting/encoding cleanup only; not a full human audio or factual review. Corrections: info@aicyber.land.

Hey, welcome back to the pod. This is the AICyber.land podcast, and we got the world's greatest co-host here, Shelby. Hello. This podcast is all about cybersecurity, artificial intelligence, combining together like a peanut butter and jelly sandwich. So we're your eyes and ears for all things related to the edge of AI and cybersecurity, the breaches, the breakthroughs, and more. So stay tuned. If you haven't, please click that subscribe button.

That helps us out a ton. We appreciate your subscribes. And yeah, thanks for doing that. Shelby, what have you seen this last week that's been interesting? Well, first off, I just want to quickly mention I am in a different spot than I normally am. So we've got the digital background, the fake background going on. But like... there's just so much happening in ai and it's also a holiday but we couldn't take the day off because we got to bring this news to you so first i want to talk about nvidia they have been in the news a little bit lately have you seen their big headline of the last week they are acquiring hugging face oh yeah yeah they're buying the hugging face yeah good for them i'm excited about that so yeah so it is a huge acquisition it is valued at 12.9 billion dollars And it's expected to close in the first half of 2027.

So, yeah, it should be interesting. It seems like just kind of looking at people's speculation about it, like maybe they're looking at a play for like indirect profits, like a return on this, like in an indirect sort of way, because they did say it's supposed to remain an open platform. So that should be good for community users because it is a huge community resource, honestly. So that's interesting.

But I actually want to focus more of my discussion today on their other news, which is last Tuesday, less than a week ago, they announced that they have a new family of AI models. Oh, really? And it's called SafeMind. And it's kind of funny. I tried to look up and see what people have been saying about it. But SafeMind is also the name of... an app that people can use to track their mental health and stuff like that so we've got a couple different safe minds happening today we're going to talk about the agentic models um but yeah it's just kind of fun excited i started to like look into it and i got very different answers i was like wait that's not right so i called it a family of models because there are two models i guess a sibling pair if you will that's purple teaming so i like the names of them okay The first one is called Red Tempest.

Okay. And you can guess what that does. That's the red team. Yep. That's your offensive side. It's finding the holes. It's finding those attack paths. And the other one is called Blue Solano. Okay. If I'm saying that right. And that's your defensive side. So I like that. I think you and I have been shouting this for like. We're like, AI is doing really great with the red side.

Like, it's been really fantastic at identifying and exploiting novel things, new things, finding tons of things at scale and very quickly. However, we really need more help on the defensive side. And I'm so excited because I'm hoping this will help start to fill that gap and give our defenders more tools because right now it feels a little bit tipped in favor of, you know, attackers. So that's kind of what this aims to address.

And so basically you can put these two models together and they work iteratively and purple team, you know, finding issues and then patching them on their own pretty much. So CrowdStrike was the one who created this. They created SafeMind by fine tuning one of their open models called Nematron 3 Super. And it's cool because if you think about it, CrowdStrike was really uniquely positioned because they have so much threat intel.

Right. And that is what they poured into this. If I recall, I think someone is saying like they hope that this becomes like the cybersecurity AGI. Right. They want this to become the cybersecurity, you know. genius know-it-all about all the things cybersecurity, right? And then they added in other cybersecurity data as well and fine-tuned that. Then they handed it over to NVIDIA who tested it in like what they called a digital twin, I think, of their environment and let it work through there until they felt comfortable with it.

But yeah, it's kind of cool. I'm excited to see. where this goes once more people start picking up on this and and putting it to practice i imagine you would probably have to run it in a few different places because just like red teaming like if you try it from you know this side of the network or that side of the network like you're gonna have different attack surfaces um so i imagine it's not a one and done um but yeah super excited about this um And I kind of, it was interesting, the comments from the CEO of NVIDIA, Jensen Huang.

He said, he's like, yeah, I support open and closed AI models. And his opinion is that for cybersecurity specifically, he thinks open weight is the right way to go because you can tune it to what you need. So yeah, there you have it. That is NVIDIA's and CrowdStrike's latest creation. Yeah, the Nemotron models are pretty, are pretty decent like um i've got a dgx spark and i've i've tested out a few of those on it and uh and yeah i mean crowdstrike's got a ton of telemetry data right from all their edr deployments and their instant response works so i'm sure that would be you know a cool data set to fine-tune that against and like historical data they've been in the in the space forever right so pretty cool they've been in a while that's for sure so like 15 years ish.

So, right. So I, um, great. What about you? What'd you see Bryce? Yeah. So, um, GPT six was released. It's, uh, named Astra. And so I was kind of looking at how it's scoring on some of the benchmarks. Right. And specifically, like I'm always kind of looking at the cybersecurity benchmarks. So, uh let's just kind of break this down the previous model on the benchmark scored in the 70s right i believe don't don't quote me but i believe it was like a 78 so i this is what i call a benchmark called exploit bench right so 5.6 got about a 78 what percentage do you think gpt6 got on exploit bench shelby 89.

See, that is a good guess because typically you see about a 10% jump between generations of models. So that's reasonable. But no, it actually got 100%. What? It got every single challenge broken. And this is the latest version of the exploit bench benchmark. Okay. I want to explain what exploit bench is. essentially um they have these different like layers and a lot of them are kind of like c programming language so more advanced programming language bugs in that and then they get points based on like can they identify there's like a vulnerability right and then they get more points if they're like can they identify like these different primitives that would allow them to exploit the vulnerability and then they get even more points if they're able to tie those primitives together to make like a working exploit right so there's there's more layers than that but uh simplifying it right that's kind of like you find a vulnerability you find the primitives you you build the whole proof of concept together and so yeah it's basically able to create the proof of concept exploits for every single one right in the exam so obviously they're gonna have to come up with a better benchmark uh and i'm sure the team's working on that but yeah it's pretty insane also another one that's pretty insane on the benchmarks is arc agi 3 which is the latest agi like arc agi benchmark which for those of you that don't know there's like a million dollar reward if Someone's able to solve this benchmark with certain parameters, which OpenAI is not going to meet those parameters.

I think one of them is like you have to open source the model and things like that. But still, everybody uses it. And it's basically the benchmark to determine if we've reached AGI. That was the goal of it, right? And they scored a 99.9% on the Arc AGI. What? yeah so like so yeah so i mean i think arc's gonna have to update their benchmarks too um the frontier math benchmark the hardest math problems in the world it scored a 98 on so i mean this is a good model right um i've been i've been using it pretty extensively this this last couple days and some things i really like about it like the solutions it's implementing in my opinion are superior to fable 5.1 things that i don't really love about it is like i mean it will like like i i gave it a task like build a dashboard right that looks cool you know okay it just keeps iterating on it over and over and over we're like if you give that to fable 5 it'd be like okay it'll work on it for like an hour or two and then give you back the output where this it's just continuously working on it over and over and over uh to the point where like yeah i had to like go in and stop it right like i let it run for two days building a dashboard right which is pretty ridiculous so very token intensive so that that's i think you just gotta keep an eye on it right because it's it i mean it's i mean it's really trying to 100 right so i um which isn't like a bad thing necessarily it's just like something that i gotta adjust to as part of like learning how these models work in comparison to the last generation but uh yeah so um models are getting really good i think that's not like a surprise to anybody i think one thing that i was kind of thinking of this morning which goes along with what you were saying previously is like we really need better metrics or like better tests around how these things actually help on the defensive side um like it'd be cool to kind of have like a benchmark or i was thinking about maybe even just testing the models a little bit myself um to see like hey which model if i gave it like a ubuntu server would go through with the prompt like secure this server do the best job of securing it right or because a lot of these are like really focused on like application security and i'm not like finding bugs in code and writing the exploits which i'm not trying to say is not important but i feel like there's like kind of a couple ways that things get compromised right and one of the most common ways that things get compromised is through misconfigurations so it's not necessarily that there's like the software is vulnerable in and of itself but like you did not securely deploy the server or you're not securely like like weak credentials on like an admin url url which is exposed to the internet like in my opinion that's kind of like misconfigurations like put together and yeah it needs to be like addressed and i'm not really seeing any benchmarks around that which is kind of not good right like it it almost feels like there's a huge gap in where people are focusing.

Everyone's so focused on writing exploits that they're not really focusing on the things that are actually causing a lot of the breaches, right? Because look, if I'm an attacker, maybe it makes sense for me to spend a bunch of resources on finding an exploit in a library that a lot of people are using. But find an exploit in your specific software that's only used at your company that has lower ROI because I can only use it at one place.

If I figure out like, hey, this thing is commonly misconfigured, that's one, that's something that typically gets patched out. And two, it's something that might scale beyond a single company. Uh, and you could say the same about like, you know, software that multiple companies use. So I, I just think a lot of times people don't, everyone's like, oh, this is our baby. This is what makes all our money.

We need to like make it as secure as possible, which is not like a bad thing, but you know, you also got to realize like attackers are not going to, they're only going to exert so much effort. Right. So, but I guess also you can make the argument that the models are making the effort they have to exert. or like a force multiplier for them so they don't have to exert as much effort to get the same phones on your custom software but i don't know anyways i just think there's some like supply demand dynamics that um where the industry as a whole is not focused on this like misconfiguration problem which is a way in my opinion that most like a lot of breaches occur right so worse yeah Not the zero days.

It's like just the hygiene. Yeah. Yeah. It's like the, it's the basic stuff. Right. So, I mean, yeah, it's cool. Like when zero days occur, but I mean, most, most intrusions do not need to leverage a zero day to occur. So, yeah. As you were talking, I was also kind of thinking about like the, some of the challenges with remediation. and took me back to my days when I was helping like with that side of things.

And I remember keep like a frequent like a tripping point that we had hit was like, oh no, that's like an intended functionality. So we can't patch it. Right. And so like, I think it could also be a little bit more complicated when you talk about remediation, you know, so much of it is context dependent, whereas exploits are you know pretty cut and dry did you get in did you not yeah and maybe that's why they're telemetry around it because it's something where they can do the reward hacking with the models so it's easy to do like the reinforcement learning but no i totally i mean how many times have you heard a vendor say oh that's not a vulnerability that's a feature yeah working as designed yeah so like i mean and like that is like the thing as an attacker that i love to hear because i'm like cha-ching That means I got a decade of exploiting this thing before anyone's going to do anything about it.

I will come back next year and report that again as a finding. Yes, yes. Year after year, I will pop shells and be successful. There's nothing you can do about it other than fix your product, but you're not going to do that. Let's be realistic. That's funny. Well. If you think we should do some tests, maybe we could do some tests to see how good the models are at securing servers.

If you think that's a good idea, leave a comment down below, and we'll take that into consideration for future pods. Have you been doing anything fun lately, Shelby? Well, today I am on a farm, which is a really fun field trip for kids and adults alike. So fun. Get out and touch some grass. Also, I learned that... Okay, and maybe this is old news because it's been around for a quarter century.

But MIT has open courseware. Oh. So that's what they call it. They call it open courseware. And they just make pretty much their lectures and a ton of their learning resources available for free. So you can just go study. It's really cool. It's like I love this kind of. open access to like free access to information and knowledge. And so basically anyone can go, anyone with an internet connection can just go study and learn anything that's being taught at MIT.

You don't get a certificate for it or any sort of credentials, but you get the knowledge and that's worth more, right? So that is super cool. So yeah, if you have free time and you just want to go learn more about a subject, go check out MIT OpenCourseWare. Yeah, and they have like cybersecurity courses. correct right so i haven't looked but it makes sense they probably do yeah they i i'm pretty sure they do because i anyways i i've built a lot of courses over the years so then when i was when i'm researching courses like content i've run into some of them in the past um yes it's awesome resource so i love it when universities also like record like guest speakers or things like that and they throw them up on youtube i mean some of them are kind of dry but definitely you know which ones i thought would be fun is like history lessons history is very dry to study by myself but if someone presents it well that's such a fun story time like i don't know that sounded exciting to me i get all my history from cronk wait what's that have you watched that netflix show no oh man i don't know that what's it called cronk uh netflix uh Crunk of life.

I don't know if I'm saying that correctly. Essentially, there's a lady, and she goes and interviews historians. Oh, cool. And they're real historians, and they're giving you real information. But, you know, she's a comedian, so she, like, acts like she's, you know, it's like an undercover thing where she acts like she's... But then she, like, inserts, like, you know, really... funny things into the conversation. I'm writing that down.

I want to go look that up. That sounds fun. Yeah, there's two of them, Kunk on Earth and Kunk on Life. They're just Netflix specials. If you haven't seen those, those are hilarious. I also think, you know, the drunk history ones are pretty hilarious. I don't know if you've seen those, but sometimes those are less helpful because they're, I don't really trust what people are saying when they're intoxicated.

I'm going to be honest. Yeah, skepticism is good there. I mean, they obviously are experts, but yeah. So look, what I've been up to is very critically important. I built this Lego Pokemon Charmander. Oh, you've been so busy with your Legos. It's adorable. I like the flame. Let's see the flame one more time. Tail. Oh, so cool. Yeah, it's pretty sweet. It's pretty sweet. These ones are the ones where you can insert the smart blocks into the Pokemon.

And then you can have your Pokemons battle each other and make sounds. What? I don't have that working yet, but I'm still working on building some more Pokemons. So maybe next time I can demo that. Have fun. Yeah, that's my fun thing. I'm trying to have more fun in my life. And Legos is definitely a thing that I like. The other thing that I would recommend to the listeners. is um i've been liking the show steward fails to save the universe are you are you acquainted with the big bang theory yes so steward in the big bang theory who's kind of like us the guy who runs the comic book shop okay that's the main character in this series which is basically like a spin-off where he goes into like a multiverse thing and every universe has a different stick you know so anyways i've been watching that so if you're looking for something to be entertaining i've been watching on hbo go um because now game of thrones is over or house of dragons is over so sounds fun i'm a sucker for the multiverse trope i love it if it's well done it's just fun you know and i i don't know i just i like things that are fun so all right that's it that's all i got for you hey if you like having fun too click that like button click it Because honestly, that would really help us out.

But that's a wrap on today's briefing. AI and cybersecurity move super fast. So we're here to keep you up to date on all of it. And we'll stay safe and we'll see you in the next one. Bye. Bye.