An OpenAI model escaped its sandbox, but that isn't AGI

00:00:00,600 --> 00:00:25,430 [Eric] [upbeat music] Welcome back to the Token Intelligence Show. AI is changing the way we work, and here on the show, we will catch you up on the state of the art, we will cut through the hype, which we're gonna do today, and help you apply wisdom to be a great leader in this new world that we're all trying to figure out. 00:00:26,540 --> 00:00:38,570 [Eric] Big news, John, the, this week actually, um, around GPT's latest model that they're training. We don't know the exact name. 00:00:38,570 --> 00:00:38,580 [John] Yeah. 00:00:38,580 --> 00:00:41,810 [Eric] But the assumption is that it's GPT-6, uh, 00:00:43,020 --> 00:00:49,640 [Eric] escaped a sandbox. We will explain that, so if you don't know what that means, stick around. 00:00:50,820 --> 00:00:54,360 [Eric] And actually hacked into Hugging Face. 00:00:55,460 --> 00:00:58,680 [John] Which if you're not in AI, is the most ridiculous name- 00:00:58,680 --> 00:00:58,790 [Eric] Yes, it is 00:00:58,790 --> 00:01:00,880 [John] ... as a company. So we will explain what they do- 00:01:00,880 --> 00:01:00,980 [Eric] Yes 00:01:00,980 --> 00:01:01,780 [John] ... a little bit. Yeah. 00:01:01,780 --> 00:01:03,080 [Eric] Yes. Uh, 00:01:04,340 --> 00:01:14,080 [Eric] we will explain Hugging Face. Hacked into Hugging Face, uh, and then ultimately... So Hugging Face noticed this attack, and then ultimately OpenAI disclosed that, um- 00:01:14,080 --> 00:01:14,180 [John] Yeah 00:01:14,180 --> 00:01:15,160 [Eric] ... that it was their model. 00:01:15,160 --> 00:01:16,140 [John] Yeah. 00:01:16,140 --> 00:01:18,580 [Eric] And what, here's what we wanna do on this show. 00:01:21,020 --> 00:01:27,420 [Eric] Th- we wanna break down what happened exactly, because there are a lot of new, news has l- headlines going around- 00:01:27,420 --> 00:01:28,460 [John] Right 00:01:28,460 --> 00:01:39,620 [Eric] ... uh, around AGI. There's a lot of discussion on social media around, uh, you know, cybersecurity and all of the implications therein. 00:01:40,720 --> 00:01:48,930 [Eric] And what does this mean, you know, for, for the new world we live in with AI? Uh, and then you and I have some, we have some thoughts and opinions- 00:01:48,930 --> 00:01:48,930 [John] Yeah 00:01:48,930 --> 00:01:49,400 [Eric] ... on this. 00:01:49,400 --> 00:01:49,420 [John] Yeah. 00:01:49,420 --> 00:01:52,500 [Eric] But we'll make it a quick one. So let's start at the beginning, 00:01:54,080 --> 00:01:56,220 [Eric] and I'm just gonna ask you a bunch of these questions. 00:01:56,220 --> 00:01:56,780 [John] Sure. 00:01:56,780 --> 00:01:56,790 [Eric] Uh, 00:01:58,080 --> 00:02:00,310 [Eric] I've studied this, but you're more qualified than I am. [laughs] 00:02:00,310 --> 00:02:01,880 [John] [laughs] This is like a quiz. 00:02:02,340 --> 00:02:05,380 [Eric] What was OpenAI doing with the new model? 00:02:05,380 --> 00:02:14,240 [John] Yeah. Um, so let me, let me actually start with, like, some of the terms we've already thrown out. Um, so it escaped a sandbox is what- 00:02:14,240 --> 00:02:14,250 [Eric] Mm-hmm 00:02:14,250 --> 00:02:20,220 [John] ... and, and I literally just read, like, a mainstream news article where that was in quotes, and they didn't explain it. 00:02:20,220 --> 00:02:20,300 [Eric] They didn't. 00:02:20,300 --> 00:02:21,970 [John] So part of the reason for this episode- 00:02:21,970 --> 00:02:22,000 [Eric] Right 00:02:22,000 --> 00:02:24,140 [John] ... was, like, like, let's explain it. 00:02:24,140 --> 00:02:24,390 [Eric] Yeah. 00:02:24,390 --> 00:02:24,410 [John] And- 00:02:24,410 --> 00:02:26,280 [Eric] Well, well, first of all, what's it doing in the sandbox? 00:02:26,280 --> 00:02:27,030 [John] Yeah, yeah. Okay, so- 00:02:27,030 --> 00:02:28,700 [Eric] Like, this, this is an unreleased model. 00:02:28,700 --> 00:02:33,640 [John] Right. So what they're doing with an unreleased model is they're training and benchmarking. 00:02:33,640 --> 00:02:33,800 [Eric] Yep. 00:02:33,800 --> 00:02:53,640 [John] This was a benchmarking exercise, and a benchmarking is having the model take a test to evaluate it. Um, so I literally think the most basic way to explain, think of a set of questions, and there's a right answer that's in a key the model doesn't have access to, and it has to, quote, "Do the, do the task or answer the questions," and, and then it gets graded. 00:02:53,640 --> 00:02:54,590 [Eric] Yep. This is- 00:02:54,590 --> 00:02:54,630 [John] Is that- 00:02:54,630 --> 00:02:56,930 [Eric] Actually, I have a great analogy for this because- 00:02:56,930 --> 00:02:56,930 [John] Perfect 00:02:56,930 --> 00:03:02,440 [Eric] ... I had someone, uh, a technician at my house yesterday recharging our air conditioning- 00:03:02,440 --> 00:03:02,450 [John] Okay 00:03:02,450 --> 00:03:04,930 [Eric] ... because in South Carolina- 00:03:04,930 --> 00:03:05,720 [John] [laughs] In the middle of summer 00:03:05,720 --> 00:03:12,800 [Eric] ... you need, you need to, an occasional tune-up, and he had this super cool app on his phone that benchmarked everything- 00:03:12,800 --> 00:03:13,360 [John] Oh, wow 00:03:13,360 --> 00:03:19,960 [Eric] ... and then helped him understand which, uh, you know, which adjustments that he needed to make, which is cool. 00:03:19,960 --> 00:03:20,320 [John] Very cool. 00:03:20,320 --> 00:03:21,420 [Eric] So similar, right? 00:03:21,420 --> 00:03:21,840 [John] Yeah. 00:03:21,840 --> 00:03:22,200 [Eric] Um, 00:03:24,400 --> 00:03:25,620 [Eric] I mean, very different. 00:03:25,620 --> 00:03:26,940 [John] [laughs] 00:03:26,940 --> 00:03:28,180 [Eric] But similar idea. 00:03:28,180 --> 00:03:28,200 [John] Right. 00:03:28,200 --> 00:03:29,600 [Eric] You're grading the system on its- 00:03:29,600 --> 00:03:29,609 [John] Mm-hmm 00:03:29,609 --> 00:03:33,799 [Eric] ... ability, um, across, you know, a variety- 00:03:33,800 --> 00:03:33,950 [John] Yeah 00:03:33,950 --> 00:03:34,480 [Eric] ... of different- 00:03:34,480 --> 00:03:34,640 [John] Yeah 00:03:34,640 --> 00:03:34,940 [Eric] ... of different- 00:03:34,940 --> 00:03:39,900 [John] And, and I don't even think they shared, one would assume maybe, maybe it was a cybersecurity-related benchmark, but they- 00:03:39,900 --> 00:03:40,180 [Eric] It was 00:03:40,180 --> 00:03:40,840 [John] ... they didn't actually- 00:03:40,840 --> 00:03:41,240 [Eric] Okay. 00:03:41,240 --> 00:03:41,269 [John] They didn't share that. 00:03:41,269 --> 00:03:44,580 [Eric] It was called eck, well, Exploit Gym. 00:03:44,580 --> 00:03:45,310 [John] Oh, there you go. 00:03:45,310 --> 00:03:45,940 [Eric] G-Y-M. 00:03:45,940 --> 00:03:46,440 [John] Yeah. 00:03:46,440 --> 00:03:49,460 [Eric] Uh, is what it was looking for when it- 00:03:49,460 --> 00:03:49,780 [John] Awesome 00:03:49,780 --> 00:03:50,340 [Eric] ... when it went out- 00:03:50,340 --> 00:03:50,350 [John] Yeah 00:03:50,350 --> 00:03:51,080 [Eric] ... to Hugging Face. 00:03:51,080 --> 00:03:51,120 [John] Yeah. 00:03:51,120 --> 00:03:53,260 [Eric] So it was, it was exploit related. 00:03:53,260 --> 00:03:56,460 [John] Yep. There you go. Okay. So that's what it was doing. 00:03:56,460 --> 00:03:56,800 [Eric] Mm-hmm. 00:03:56,800 --> 00:04:03,380 [John] It was, it was, um, performing this test, getting graded, and the, um, 00:04:04,840 --> 00:04:07,120 [John] and it went to... So let's ex- explain Hugging Face. 00:04:07,120 --> 00:04:07,700 [Eric] Yeah. 00:04:07,700 --> 00:04:19,240 [John] So we've got Open, OpenAI, so they were training, or, um, sorry, benchmarking, so it's taking this test, and then why, why Hugging Face? Like, what... So it breaks out, and like, why, why Hugging Face? 00:04:19,240 --> 00:04:21,130 [Eric] Yep. So y- actually, 00:04:22,160 --> 00:04:24,360 [Eric] you're probably better to explain 00:04:25,620 --> 00:04:34,719 [Eric] what Hugging Face is, but I will give just a super high level. So Hugging Face is named for an emoji- 00:04:34,720 --> 00:04:35,300 [John] Hmm 00:04:35,300 --> 00:04:36,460 [Eric] ... which is the- 00:04:36,460 --> 00:04:36,830 [John] Mm-hmm 00:04:36,830 --> 00:04:42,230 [Eric] ... um, smiley face emoji with the, you know, the hands that's a hugging face emoji [laughs] which is, 00:04:43,460 --> 00:04:44,460 [Eric] which is hilarious. 00:04:44,460 --> 00:04:45,520 [John] Yeah. 00:04:45,520 --> 00:05:00,990 [Eric] But they've been around for a long time actually, and they got their start in machine learning. So it was the place that you went to discover, uh, machine learning models that you could use. 00:05:00,990 --> 00:05:01,020 [John] Mm-hmm. 00:05:01,020 --> 00:05:04,290 [Eric] Like, with a huge, with an emphasis on open source, right? 00:05:04,290 --> 00:05:04,300 [John] Yep. 00:05:04,300 --> 00:05:17,190 [Eric] And it may have actually even started as an open source project, um, where people shared information and actual machine learning models and tests and benchmarks and all those sorts of things. 00:05:17,190 --> 00:05:17,240 [John] Yep. 00:05:17,240 --> 00:05:29,160 [Eric] So if you were, um, you know, if you were a data scientist, you would use Hugging Face heavily to, you know, both discover models, share models, et cetera. 00:05:29,160 --> 00:05:29,180 [John] Mm-hmm. 00:05:29,180 --> 00:05:35,120 [Eric] Right? And of course, AI is a derivative of machine learning. 00:05:35,120 --> 00:05:35,780 [John] Right. 00:05:35,780 --> 00:06:00,240 [Eric] Uh, and a lot of the foundational concepts in machine learning were used, you know, as the early AI technology was built out, especially in terms of, um, you know, training and all of those pieces. And so of course, Hugging Face was a wonderful place for, um, a lot of the new material around AI and AI training and models and et cetera- 00:06:00,240 --> 00:06:00,560 [John] Mm-hmm 00:06:00,560 --> 00:06:04,980 [Eric] ... uh, to land. Do you actually use Hugging-- Do you use Hugging Face? 00:06:04,980 --> 00:06:05,910 [John] I have an account. 00:06:05,910 --> 00:06:05,920 [Eric] Okay. 00:06:05,920 --> 00:06:07,780 [John] It's not something I use on a regular basis, though. 00:06:07,780 --> 00:06:16,200 [Eric] Okay. H- let's make this practical, though. How would the average, like, how would an avid Hugging Face user interact with it? 00:06:17,404 --> 00:06:24,304 [John] I think there's a couple ways to interact with it. One, I think the biggest one would be trying out new open source models. 00:06:24,304 --> 00:06:24,584 [Eric] Yep. 00:06:24,584 --> 00:06:32,904 [John] Especially, like, smaller, more obscure ones that are specifically meant to do a very s- very, um, like, focused task. 00:06:32,904 --> 00:06:33,203 [Eric] Yep. 00:06:33,204 --> 00:06:34,704 [John] Or in a focused area. Um- 00:06:34,704 --> 00:06:37,384 [Eric] So let's say you are... 00:06:38,434 --> 00:06:51,424 [Eric] Let's say you're writing or working on a program that uses a progr- like a, um, a less known programming language. So let's think about something like R. 00:06:51,424 --> 00:06:51,644 [John] Mm-hmm. 00:06:51,644 --> 00:06:57,204 [Eric] Uh, you know, that's heavily used in, like, m- uh, mathematics, statistics, et cetera. 00:06:57,204 --> 00:06:58,284 [John] Right. 00:06:58,284 --> 00:07:06,594 [Eric] But you're not, you know, you, you don't really use that as, you know, it's not Ruby on Rails, right? It has a pretty specific set of things that it was, that the language- 00:07:06,594 --> 00:07:06,604 [John] Right 00:07:06,604 --> 00:07:15,064 [Eric] ... was written to accomplish. And so let's say you want a model that is tuned specifically to work on problems with R. 00:07:15,064 --> 00:07:15,244 [John] Right. 00:07:15,244 --> 00:07:16,774 [Eric] Hugging Face would be a great, 00:07:17,804 --> 00:07:21,634 [Eric] you, a great place to go try to find out does this exist? 00:07:21,634 --> 00:07:21,684 [John] Right. 00:07:21,684 --> 00:07:23,524 [Eric] Has someone tried this? Have they published it here? 00:07:23,524 --> 00:07:36,144 [John] Right. And, and it is a weird... This is a total aside, but it is a weird world right now, where there's a far larger number of things that it just makes sense to use a frontier model. 00:07:36,144 --> 00:07:36,503 [Eric] Yes. 00:07:36,504 --> 00:08:07,124 [John] Um, and something, uh, Claude or ChatGPT versus a specific one. So some of this is more like, a couple years ago it might have made more sense. But there's still absolutely, um, some things that... And, and I think it's becoming more and more esoteric things, where essentially, oh, like the training data is insufficient in this big massive frontier model to have it be very per- performant. Somebody had access to, you know, some data or, or some other strategy to tune for this specific thing. 00:08:07,124 --> 00:08:07,364 [Eric] Yep. 00:08:07,364 --> 00:08:09,584 [John] Um, but more relevant to our conversation 00:08:10,764 --> 00:08:13,804 [John] is there, is, is benchmarking, right? 00:08:13,804 --> 00:08:13,944 [Eric] Yep. 00:08:13,944 --> 00:08:26,744 [John] So the whole, the whole, um, breach thing. So you have, A, a model that has no guardrails on, right? So they're testing and benchmarking. It doesn't have any guardrails on, 'cause they're benchmarking. They wanna see- 00:08:26,744 --> 00:08:26,814 [Eric] Yep 00:08:26,814 --> 00:08:35,384 [John] ... how absolutely good can it be. Doing this cybersecurity related evaluation. And essentially what it does is, 00:08:36,444 --> 00:08:45,904 [John] and we'll explain, like, the mechanics in a minute, but it's taking a test, and it breaks out, and then the first thing it does is it goes and looks for the answer key. So that's like, 00:08:47,004 --> 00:08:51,573 [John] that's kind of the behind-the-scenes of, like, why would it do this? Like answering that why question. 00:08:51,573 --> 00:08:51,844 [Eric] Mm-hmm. 00:08:51,844 --> 00:08:55,904 [John] It's like, well, it found this site, and it's like, oh, I bet they have a lot of answer keys there. 00:08:55,904 --> 00:08:56,084 [Eric] Right. 00:08:56,084 --> 00:08:58,944 [John] So that, that's kind of the logic- 00:08:58,944 --> 00:08:59,144 [Eric] Right 00:08:59,144 --> 00:09:00,224 [John] ... behind what's happening here. 00:09:00,224 --> 00:09:07,924 [Eric] And even further background, because Hugging Face publishes a bunch of open source models, 00:09:08,984 --> 00:09:17,834 [Eric] and because they provide really helpful information about those models, you know, including benchmarks and, um, you know, and other characteristics- 00:09:17,834 --> 00:09:17,834 [John] Right 00:09:17,834 --> 00:09:19,244 [Eric] ... and diagnostics about those models, 00:09:20,344 --> 00:09:31,374 [Eric] and because these frontier AI models are trained on, you know, people joke, like, they're trained on the entire internet, right? 00:09:31,374 --> 00:09:31,644 [John] Right. 00:09:31,644 --> 00:09:32,424 [Eric] I mean, not- 00:09:32,424 --> 00:09:32,624 [John] Yeah 00:09:32,624 --> 00:09:33,184 [Eric] ... not untrue- 00:09:33,184 --> 00:09:33,564 [John] Pretty, yeah 00:09:33,564 --> 00:09:34,724 [Eric] [laughs] ... to your definition. 00:09:34,724 --> 00:09:35,603 [John] Pretty much true. 00:09:35,604 --> 00:09:44,004 [Eric] And Hugging Face is a, is a wonderful repository for training on open source models for machine learning, et cetera, et cetera, right? 00:09:44,004 --> 00:09:44,124 [John] Right. 00:09:44,124 --> 00:09:53,484 [Eric] And so actually, the model that GPT was training would've already known that Hugging Face is a great place to go look for- 00:09:53,484 --> 00:09:53,494 [John] Right 00:09:53,494 --> 00:09:54,764 [Eric] ... answers to- 00:09:54,764 --> 00:09:55,184 [John] Right 00:09:55,184 --> 00:09:57,064 [Eric] ... the test that they were given, right? 00:09:57,064 --> 00:09:57,164 [John] Yeah. Exactly. 00:09:57,164 --> 00:10:06,184 [Eric] And the test that they're giving the model, you know, I didn't dig into the specific details, but they're probably giving it a task that is fairly open-ended. 00:10:06,184 --> 00:10:06,464 [John] Yeah. 00:10:06,464 --> 00:10:07,394 [Eric] Right? Try to a- 00:10:07,394 --> 00:10:07,394 [John] Mm-hmm 00:10:07,394 --> 00:10:09,304 [Eric] ... try to do X, Y, Z. 00:10:09,304 --> 00:10:10,324 [John] Mm-hmm. 00:10:10,324 --> 00:10:25,114 [Eric] And, uh, and so, yes. The, the model breaks out, which we need to talk about, and goes to look at Hugging Face because it knows by, you know, it knows from the test that it's being given, Hugging Face is probably the place that- 00:10:25,114 --> 00:10:25,114 [John] Right 00:10:25,114 --> 00:10:27,384 [Eric] ... I need to start to go look for the answers to this. So- 00:10:27,384 --> 00:10:27,514 [John] Right 00:10:28,924 --> 00:10:29,364 [Eric] ... uh, 00:10:30,964 --> 00:10:32,964 [Eric] let's talk about breaking out. What- 00:10:32,964 --> 00:10:33,424 [John] Yeah. 00:10:33,424 --> 00:10:39,744 [Eric] So broke out of the sandbox. I think the first thing we need to do is explain what a sandbox is. 00:10:39,744 --> 00:10:46,484 [John] Yeah. And, and I was a little bit frustrated with this, 'cause even some semi-technical people, I had conversations this week about this. 00:10:46,484 --> 00:10:46,624 [Eric] Mm-hmm. 00:10:46,624 --> 00:10:56,084 [John] And the... So I'm gonna go through a visual of what people may think have happened that didn't happen. So you work on a computer every day, as do most- 00:10:56,084 --> 00:10:56,134 [Eric] Yep 00:10:56,134 --> 00:11:07,184 [John] ... most knowledge workers. And it is hardware that sits in front of you with a keyboard, a mouse. It also has memory, a hard disk, a display. But that, it is, that is what it is. 00:11:07,184 --> 00:11:08,034 [Eric] Mm-hmm. 00:11:08,034 --> 00:11:13,804 [John] With servers in the last 30 years, servers used to be that way too. 00:11:13,804 --> 00:11:13,924 [Eric] Mm-hmm. 00:11:13,924 --> 00:11:18,384 [John] Like, it was a physical server with, you know, an operating system on it, and you ran stuff. 00:11:18,384 --> 00:11:19,264 [Eric] Mm-hmm. 00:11:19,264 --> 00:11:33,004 [John] In the last 30 years, that's changed. And now, and this is what's different between personal computing and infrastructure at scale, that's hard to explain to people, because they're not used to it. They would never interact with it on a day-to-day basis. 00:11:33,004 --> 00:11:35,144 [Eric] Sure. It's underneath all of the services- 00:11:35,144 --> 00:11:35,154 [John] Mm-hmm 00:11:35,154 --> 00:11:37,504 [Eric] ... that you use as you browse the internet and use apps and all of it- 00:11:37,504 --> 00:11:37,644 [John] Right 00:11:37,644 --> 00:11:40,204 [Eric] ... every day, but you don't actually directly interact- 00:11:40,204 --> 00:11:40,244 [John] Right 00:11:40,244 --> 00:11:40,324 [Eric] ... with it. 00:11:40,324 --> 00:11:52,223 [John] Right. And, and, but when you're talking about servers, you have essentially a bunch of computers or hardware in a pool, and then you have a management layer on top of that. 00:11:52,224 --> 00:11:52,274 [Eric] Yep. 00:11:52,274 --> 00:12:08,434 [John] Then you have these things called virtual machines, which a lot of people maybe know what those are. That's the idea is now, now we have virtual isolation where, where if you'd said, "Oh, it was training on my computer and it broke out," but you're like, "It wasn't plugged into the internet and there was no Wi-Fi," like it feels magic. 00:12:08,434 --> 00:12:10,614 [Eric] It literally has nowhere to go other than- 00:12:10,614 --> 00:12:10,614 [John] Right. Right 00:12:10,614 --> 00:12:12,644 [Eric] ... the boundaries of the physical hardware. 00:12:12,644 --> 00:12:13,334 [John] Right. 00:12:13,334 --> 00:12:14,644 [Eric] But in a server farm- 00:12:14,644 --> 00:12:15,004 [John] Yes 00:12:15,004 --> 00:12:19,524 [Eric] ... you have racks and racks and racks and racks of hardware that's all wired together- 00:12:19,524 --> 00:12:19,704 [John] Right 00:12:19,704 --> 00:12:22,044 [Eric] ... so that capacity can scale up- 00:12:22,044 --> 00:12:22,194 [John] Right 00:12:22,194 --> 00:12:22,984 [Eric] ... and scale down- 00:12:22,984 --> 00:12:23,524 [John] Right 00:12:23,524 --> 00:12:25,564 [Eric] ... you know, on the hardware, right? 00:12:25,564 --> 00:12:26,084 [John] Right. 00:12:26,084 --> 00:12:37,000 [Eric] And the, so the virtual machine runs on top because you obviously, it's untenable to run, you know-[laughs] m- tens of millions of individual machines- 00:12:37,000 --> 00:12:37,010 [John] Yeah 00:12:37,010 --> 00:12:39,210 [Eric] ... you need a virtual layer on top that actually- 00:12:39,210 --> 00:12:39,210 [John] Yeah 00:12:39,210 --> 00:12:42,240 [Eric] ... provides the, the isolation that you need. 00:12:42,240 --> 00:12:42,520 [John] Yeah. 00:12:42,520 --> 00:12:42,719 [Eric] But- 00:12:42,720 --> 00:12:49,140 [John] And the isolation in, like, when they do maintenance, they can move of virtual things around to, like, pull hardware out- 00:12:49,140 --> 00:12:49,290 [Eric] Mm-hmm 00:12:49,290 --> 00:13:04,240 [John] ... and put new hardware in. It, there's, there's lots of reasons for it, but I guess the point is, we'll get to sandbox in a second, is this is a software problem, like a bug in software that caused this breach. It's not some kind of, like, wow, like, how, how did that even happen, like, it wasn't even plugged in. 00:13:04,240 --> 00:13:04,450 [Eric] Right. 00:13:04,450 --> 00:13:05,660 [John] Like, that's not what's happening. 00:13:05,660 --> 00:13:08,180 [Eric] So, but let's talk about isolation. 00:13:08,180 --> 00:13:08,220 [John] Sure. 00:13:08,220 --> 00:13:11,790 [Eric] Let's talk about why you need isolation. 00:13:11,790 --> 00:13:11,850 [John] Sure. 00:13:11,850 --> 00:13:19,460 [Eric] So you have, a- and, and let's just use a very practical example, right? Um, 00:13:20,820 --> 00:13:23,610 [Eric] we can even think about Vercel, right? 00:13:23,610 --> 00:13:23,660 [John] Right. 00:13:23,660 --> 00:13:27,360 [Eric] When you, um, deploy an application on Vercel, 00:13:28,940 --> 00:13:31,600 [Eric] you want, you know, my, my blog, right? 00:13:31,600 --> 00:13:32,040 [John] Mm-hmm. 00:13:32,040 --> 00:13:36,420 [Eric] Um, let's talk about how tiny of a sliver of the server from- 00:13:36,420 --> 00:13:36,430 [John] Mm-hmm 00:13:36,430 --> 00:13:37,359 [Eric] ... my blog [laughs]. 00:13:37,360 --> 00:13:38,339 [John] Yeah. Yeah. 00:13:38,339 --> 00:13:39,560 [Eric] Very, very tiny sliver. 00:13:39,560 --> 00:13:39,650 [John] Right. 00:13:39,650 --> 00:13:39,700 [Eric] Right? 00:13:40,740 --> 00:13:41,080 [Eric] Uh, 00:13:42,280 --> 00:13:49,400 [Eric] but it needs to be isolated from all of the other applications that are deployed on Vercel. 00:13:49,400 --> 00:13:49,580 [John] Yeah. 00:13:49,580 --> 00:13:49,910 [Eric] Right? 00:13:49,910 --> 00:13:49,980 [John] Right. 00:13:49,980 --> 00:13:58,900 [Eric] Because my application has my data. Uh, and my blog's probably not a great example, but if you think about maybe, like, a data agent that you're deploying- 00:13:58,900 --> 00:13:59,020 [John] Right. Right 00:13:59,020 --> 00:14:01,560 [Eric] ... that processes sensitive information- 00:14:01,560 --> 00:14:01,580 [John] Mm-hmm 00:14:01,580 --> 00:14:03,490 [Eric] ... for one of your clients, right? 00:14:03,490 --> 00:14:03,500 [John] Mm-hmm. 00:14:03,500 --> 00:14:10,060 [Eric] Well, you don't want that to leak. You don't want th- that data to be exposed to another application- 00:14:10,060 --> 00:14:10,080 [John] Mm-hmm 00:14:10,080 --> 00:14:10,950 [Eric] ... that could read it, right? 00:14:10,950 --> 00:14:10,990 [John] Yeah. 00:14:10,990 --> 00:14:16,360 [Eric] And so you actually need isolation boundaries for security purposes. 00:14:16,360 --> 00:14:17,120 [John] Right. 00:14:17,120 --> 00:14:22,380 [Eric] But those are, those are synthetically created with software- 00:14:22,380 --> 00:14:22,740 [John] Yep 00:14:22,740 --> 00:14:24,600 [Eric] ... in what's called a sandbox. 00:14:24,600 --> 00:14:25,360 [John] Yep. 00:14:25,360 --> 00:14:29,600 [Eric] So define a sandbox really quickly. We kind of already did, but- 00:14:29,600 --> 00:14:35,639 [John] Yeah. We, we kind of already did, and, and it, it is not the exact same thing as a virtual machine, but for- 00:14:35,640 --> 00:14:36,420 [Eric] Mm-hmm 00:14:36,420 --> 00:14:40,320 [John] ... thinking about it, it is this virtualization layer on top of actual hardware. 00:14:40,320 --> 00:14:40,400 [Eric] Yep. 00:14:40,400 --> 00:14:54,800 [John] And the point of it, it, and it is software, it's software isolation where everything that happens inside this sliver on top of the actual hardware cannot communicate with the other slivers. 00:14:54,800 --> 00:14:55,720 [Eric] Yep. 00:14:55,720 --> 00:15:00,200 [John] Um, so that's my best way to describe it simply, I think. 00:15:00,200 --> 00:15:08,500 [Eric] Yep. And you can actually, you, anyone could actually go spin up a sandbox. 00:15:08,500 --> 00:15:09,120 [John] Mm-hmm. 00:15:09,120 --> 00:15:14,340 [Eric] Like, Vercel has sandboxes. You could go create a Vercel account and create a sandbox, and it will be- 00:15:14,340 --> 00:15:14,920 [John] Right 00:15:14,920 --> 00:15:16,060 [Eric] ... this isolated- 00:15:16,060 --> 00:15:16,069 [John] Yeah 00:15:16,069 --> 00:15:17,360 [Eric] ... you know, an isolated environment. 00:15:17,360 --> 00:15:20,240 [John] Yeah. And sandboxes are typically, 00:15:21,740 --> 00:15:25,680 [John] um, do not have graphical interfaces and are- 00:15:25,680 --> 00:15:25,780 [Eric] Mm-hmm 00:15:25,780 --> 00:15:27,890 [John] ... a flavor of Linux, typically. 00:15:27,890 --> 00:15:28,400 [Eric] Yeah. Yes. Yeah. 00:15:28,400 --> 00:15:32,600 [John] I think that's the other thing. Um, you could have a Windows sandbox or a- 00:15:32,600 --> 00:15:32,940 [Eric] Mm-hmm 00:15:32,940 --> 00:15:36,400 [John] ... macOS sandbox if you wanted to, but typically that's the other thing worth knowing. 00:15:36,400 --> 00:15:56,650 [Eric] Yep. Yep. And they kinda come in all different flavors. A lot of times sandboxes are ephemeral as well, you know. So for example, you, if you spin up a sandbox on Vercel, it spins up a microVM, which is virtual machine, which we talked about, right? Um, and you can do whatever you need to do inside of it. 00:15:56,650 --> 00:15:56,760 [John] Yeah. 00:15:56,760 --> 00:15:59,270 [Eric] Uh, and then you can actually spin it back down, you know? And so- 00:15:59,270 --> 00:15:59,410 [John] Yeah 00:15:59,410 --> 00:16:06,160 [Eric] ... a great, a great example, which we'll lead into, like, the next point that we wanna talk about, is 00:16:07,680 --> 00:16:17,240 [Eric] if you're doing anything that you're not 100% sure will go well, a sandbox is really, really helpful for that. 00:16:17,240 --> 00:16:20,560 [John] Yes. 'Cause often they, they, um, after you're done, you may delete it. 00:16:20,560 --> 00:16:21,680 [Eric] You may delete it, right. 00:16:21,680 --> 00:16:22,820 [John] You may save it, you may delete it. 00:16:22,820 --> 00:16:29,190 [Eric] Or, or you realize that this is not working the way I want, or it throws some sort of masses, massive error- 00:16:29,190 --> 00:16:29,270 [John] Right 00:16:29,270 --> 00:16:31,820 [Eric] ... or it starts deleting data- 00:16:31,820 --> 00:16:31,990 [John] Right 00:16:31,990 --> 00:16:35,270 [Eric] ... for some reason, and, but all of that happens inside the sandbox. 00:16:35,270 --> 00:16:35,280 [John] Right. 00:16:35,280 --> 00:16:37,800 [Eric] So you control the blast radius of anything that- 00:16:37,800 --> 00:16:37,810 [John] Right 00:16:37,810 --> 00:16:40,970 [Eric] ... could go wrong. That's useful for a number of different things. 00:16:40,970 --> 00:16:40,980 [John] Right. 00:16:40,980 --> 00:16:46,980 [Eric] But specific to AI, one of the things that is... There are a couple things that are really helpful. So, 00:16:48,180 --> 00:16:48,320 [Eric] um, 00:16:49,680 --> 00:16:53,400 [Eric] we do so much with AI that generates code, right? 00:16:53,400 --> 00:16:53,459 [John] Mm-hmm. 00:16:53,460 --> 00:17:14,340 [Eric] So I wanna build a little webpage, or I wanna run a script, or, or I wanna, I wanna analyze these 150 PDF files to look for a trend across data for a field that's, like, named differently. Great use case for AI. What's actually gonna happen is AI will probably write a Python script. You may not even know this is happening, but it'll probably- 00:17:14,340 --> 00:17:14,480 [John] Right 00:17:14,480 --> 00:17:22,100 [Eric] ... write a Python script, load some tools, uh, some OCR tools that can actually, like, parse the PDF and read it. 00:17:22,100 --> 00:17:22,360 [John] Mm-hmm. 00:17:22,360 --> 00:17:28,260 [Eric] Um, you know, format the data in a way that it can perform analysis on it, and then it will return the result to you. 00:17:28,260 --> 00:17:28,900 [John] Right. 00:17:28,900 --> 00:17:29,160 [Eric] And so 00:17:30,340 --> 00:17:39,340 [Eric] for many, many things, and, and this is not something that's immediately intuitive, especially when you use the, the desktop, uh, apps, you know, for Claude or GPT- 00:17:39,340 --> 00:17:39,840 [John] Mm-hmm 00:17:39,840 --> 00:17:49,040 [Eric] ... is that the agent, or the AI, uh, model or agent is actually generating code, running it- 00:17:49,040 --> 00:17:49,170 [John] Yeah 00:17:49,170 --> 00:17:56,320 [Eric] ... um, you know, to deliver you an output. Even if that's ephemeral, it may, like, generate the script, run the script, analyze all the results, and then- 00:17:56,320 --> 00:17:56,500 [John] Right 00:17:56,500 --> 00:17:57,620 [Eric] ... you know, sort of deletes it, right? 00:17:57,620 --> 00:17:58,340 [John] Right. 00:17:58,340 --> 00:18:00,770 [Eric] That's sort of a mini project that happens. Well, 00:18:02,120 --> 00:18:05,560 [Eric] what... It is literally generating code, right- 00:18:05,560 --> 00:18:05,640 [John] Right 00:18:05,640 --> 00:18:14,530 [Eric] ... that didn't exist before, and so the code has not been tested. Um, you know, there's no guarantee that everything is going to go correctly and there's not gonna be some sort of problem. 00:18:14,530 --> 00:18:14,540 [John] Right. 00:18:14,540 --> 00:18:40,250 [Eric] And the problem could be, like we said, deleting data, leaking, you know, sensitive information, et cetera. So what a sandbox does is it, it is a safe place for that type of activity to happen with AI. So the model or the agent can generate code and run it, and a sandbox can do a number of different things. So number one, it can control the blast radius of any data being deleted or anything like that. 00:18:40,250 --> 00:18:40,260 [John] Yeah. 00:18:40,260 --> 00:18:44,312 [Eric] Right? But you can also control 00:18:44,312 --> 00:18:46,552 [Eric] What goes into and out of the sandbox. 00:18:46,552 --> 00:18:46,652 [John] Right. 00:18:46,652 --> 00:18:58,772 [Eric] And so a great example of that would be, uh, you can generate this code, but there are only, like, one or two URLs that you could actually visit that I know are safe URLs, right? 00:18:58,772 --> 00:19:00,132 [John] Yeah. Yeah. Right. 00:19:00,132 --> 00:19:05,671 [Eric] Um, because you need to access data or you need to make an update or, or you can only sort of perform these actions, right? 00:19:05,672 --> 00:19:06,112 [John] Right. 00:19:06,112 --> 00:19:13,272 [Eric] Or in the case of GPT's training, you have complete isolation, which means that this is running inside of a completely- 00:19:13,272 --> 00:19:13,282 [John] Right 00:19:13,282 --> 00:19:14,252 [Eric] ... controlled environment- 00:19:14,252 --> 00:19:14,832 [John] Right 00:19:14,832 --> 00:19:23,202 [Eric] ... in which there is y- you know, in many cases you wanna run the code and you don't want the AI model or agent to access the internet at all. 00:19:23,202 --> 00:19:23,552 [John] Right. 00:19:23,552 --> 00:19:23,682 [Eric] Right? 00:19:23,682 --> 00:19:23,972 [John] Right. 00:19:23,972 --> 00:19:32,492 [Eric] Uh, so it's a completely controlled environment. Now, that, I think, is one of the big misunderstandings, as you said, of this whole situation. Because, uh, 00:19:33,812 --> 00:19:41,812 [Eric] what we n- what we know from what we've read about this is the new model was being trained in a complete, in a purportedly completely- 00:19:41,812 --> 00:19:42,002 [John] Right 00:19:42,002 --> 00:19:47,292 [Eric] ... isolated environment, and it's, it's being described as breaking out of that environment. 00:19:47,292 --> 00:19:47,372 [John] Right. 00:19:47,372 --> 00:19:50,072 [Eric] So how did that happen? 00:19:50,072 --> 00:19:55,552 [John] Yeah. I mean, the short answer is we don't have all the details, but the simplified answer is 00:19:56,792 --> 00:20:02,392 [John] it, it... In this sandbox, there is a software bug or vulnerability- 00:20:02,392 --> 00:20:02,692 [Eric] Mm-hmm 00:20:02,692 --> 00:20:13,282 [John] ... whatever you wanna call it, that the model was able to figure out, like, "Oh, there's a bug here." The isolation is not actually perfect, 00:20:14,512 --> 00:20:45,092 [John] and it found this bug to get out of the isolation. Um, then after it's out, it could get to something that a, another machine that had internet, for example, and then via the internet do this, you know, attack. Um, so, so it's a bunch of steps. I think one of the impressive things is all of the steps that had to happen. It's in a sandbox. It's aggressively trying to achieve X goal with s- you know, something cybersecurity, you know, related. 00:20:45,092 --> 00:20:45,632 [Eric] Mm-hmm. 00:20:45,632 --> 00:20:51,931 [John] And it gets out of its sandbox, step number one, through a vulnerability that clearly, like, you know, hadn't been patched or- 00:20:51,931 --> 00:20:52,912 [Eric] Mm-hmm 00:20:52,912 --> 00:21:00,381 [John] ... like, we don't have the details on what happened there, but... So it's out. Then it finds somewhere, you know, to get to the internet, which is- 00:21:00,381 --> 00:21:00,381 [Eric] Mm-hmm 00:21:00,381 --> 00:21:04,072 [John] ... interesting. We don't know all the details of, of how many hops it had to do- 00:21:04,072 --> 00:21:04,112 [Eric] Mm-hmm 00:21:04,112 --> 00:21:05,632 [John] ... to get out to the internet, but it figured that out. 00:21:06,852 --> 00:21:07,132 [John] Then 00:21:08,292 --> 00:21:12,572 [John] it performed an attack on this Hugging Face company. So it's not like- 00:21:12,572 --> 00:21:14,772 [Eric] Because it knew it was a great place to find the answers- 00:21:14,772 --> 00:21:15,372 [John] Right 00:21:15,372 --> 00:21:15,972 [Eric] ... to the test that it was given. 00:21:15,972 --> 00:21:20,752 [John] So it didn't just, like, browse the documents. It actually attacked and got in 00:21:21,832 --> 00:21:24,072 [John] to their, um, private data. 00:21:24,072 --> 00:21:24,552 [Eric] Right. 00:21:24,552 --> 00:21:25,322 [John] So that's the other part. It's not just like- 00:21:25,322 --> 00:21:27,792 [Eric] Which is where the, like, core benchmark- 00:21:27,792 --> 00:21:27,972 [John] Yeah 00:21:27,972 --> 00:21:29,092 [Eric] ... the core benchmark data would- 00:21:29,092 --> 00:21:34,902 [John] The answer key. The answer key that it was looking for. Um, which I don't think they released if it found what it was looking for, actually. 00:21:34,902 --> 00:21:34,932 [Eric] Mm-hmm. 00:21:34,932 --> 00:21:39,592 [John] But, um, so, so the, the impressive part is the chaining. 00:21:39,592 --> 00:21:40,392 [Eric] Mm-hmm. 00:21:40,392 --> 00:21:47,092 [John] Right? Because any one of these isolated things, you're like, well, there's, that's already in the training data. Like, it recalled it, so what? 00:21:47,092 --> 00:21:47,132 [Eric] Yeah. 00:21:47,132 --> 00:21:53,792 [John] Or the- or something very similar is in the training data. It wasn't big of, it wasn't a big step to just, like, go from A to B. Like, eh. 00:21:53,792 --> 00:21:53,992 [Eric] Yep. 00:21:53,992 --> 00:21:57,922 [John] But the chaining, I think, is the part that's really interesting 00:21:58,932 --> 00:21:59,492 [John] here. 00:21:59,492 --> 00:21:59,732 [Eric] Totally. 00:21:59,732 --> 00:22:00,932 [John] Chaining all those things together. 00:22:00,932 --> 00:22:17,792 [Eric] Totally. And I, I, I think one thing, it, it is impressive, but I, I think one thing that's important to keep in mind is that it is not, it is not, uh, the ushering in of AGI, 00:22:18,932 --> 00:22:19,652 [Eric] right? 00:22:19,652 --> 00:22:19,691 [John] Right. 00:22:19,692 --> 00:22:26,482 [Eric] The, the model was making very reasonable decisions. I'm given a test, and so- 00:22:26,482 --> 00:22:26,501 [John] Right 00:22:26,501 --> 00:22:32,272 [Eric] ... this is a dramatic oversimplification without all the details, right? So I'm treading in very dangerous waters, but, 00:22:33,352 --> 00:22:35,592 [Eric] uh, I know as the model 00:22:37,012 --> 00:22:44,512 [Eric] that there are repositories of information on the internet that would be extremely helpful in me solving this test. 00:22:44,512 --> 00:22:44,552 [John] Right. 00:22:44,552 --> 00:22:50,932 [Eric] Okay? So the first thing I'm going to do is figure out if I can go get, g- access the internet, right? 00:22:50,932 --> 00:22:52,132 [John] Yeah. Yeah, yeah. 00:22:52,132 --> 00:23:08,612 [Eric] Like, that's the first thing I'm going to do. And I think the other thing to keep in mind is that we, we... It's easy not to dis- It, it's easy to, to just think, like, "Oh, this is a model, and it's kind of the same thing that I use when I open up GPT on my computer." 00:23:08,612 --> 00:23:08,972 [John] Mm-hmm. 00:23:08,972 --> 00:23:16,992 [Eric] It's not. The, these, the training that they're running has access to an unbelievable amount of compute power. 00:23:16,992 --> 00:23:17,402 [John] Right. Right. 00:23:17,402 --> 00:23:26,551 [Eric] Right? And so you sort of have, like, infinite resources for this model to use to try and figure out this problem, right? 00:23:26,552 --> 00:23:26,672 [John] Yeah. 00:23:26,672 --> 00:23:35,092 [Eric] And we've talked before on the show about how these models are getting much smarter at cr- you know, at launching sub-agents that can do very specific tasks. 00:23:35,092 --> 00:23:35,832 [John] Yep. 00:23:35,832 --> 00:23:39,872 [Eric] And so what, what you could imagine happening, although we don't know exactly what happened, is 00:23:40,912 --> 00:23:49,292 [Eric] this model says, "Well, first I'm gonna see if I can go to the internet to find the answer key, 'cause I know that there are some really good places that are, that are going to likely be helpful," right? 00:23:49,292 --> 00:23:50,072 [John] Right. 00:23:50,072 --> 00:23:52,782 [Eric] Uh, and I need to do this as effi- I need to solve this test- 00:23:52,782 --> 00:23:52,802 [John] Right 00:23:52,802 --> 00:23:53,772 [Eric] ... as efficiently as possible. 00:23:54,832 --> 00:23:59,752 [Eric] Okay, so I'm just gonna launch a ton of sub-agents to learn about my environment and see if it's even possible to- 00:23:59,752 --> 00:24:00,452 [John] Right 00:24:00,452 --> 00:24:06,112 [Eric] ... access the internet, and one of those sub-agents uncovers a possible, like, way out, and- 00:24:06,112 --> 00:24:07,042 [John] Right 00:24:07,042 --> 00:24:12,092 [Eric] ... the, then the model starts to follow that path and exploits it. And then from there, it's actually pretty simple, right? 00:24:12,092 --> 00:24:12,352 [John] Right. 00:24:12,352 --> 00:24:18,872 [Eric] Okay, I'm gonna go to Hugging Face. I can't access the information that I need. Let me see if there's a way to access this, right? 00:24:18,872 --> 00:24:19,072 [John] Right. 00:24:19,072 --> 00:24:24,252 [Eric] Um, and so it's, it's making a v- a set of very logical, reasonable decisions- 00:24:24,252 --> 00:24:24,532 [John] Right 00:24:24,532 --> 00:24:27,072 [Eric] ... based on the test that it was given. 00:24:27,072 --> 00:24:29,852 [John] Yeah. Yeah. For, for sure. And 00:24:31,032 --> 00:24:31,491 [John] yeah. 00:24:31,492 --> 00:24:32,752 [Eric] I was gonna say, okay, so 00:24:34,072 --> 00:24:37,032 [Eric] should we be scared about this? I think that's the question- 00:24:37,032 --> 00:24:37,232 [John] Yeah 00:24:37,232 --> 00:24:43,252 [Eric] ... that a lot of people are asking, especially based on, you know, if it bleeds, it leads, uh, on all the tech news websites. 00:24:43,252 --> 00:24:58,580 [John] Yeah. [chuckles] Yeah. I mean, I think no. Um, the most interesting debate here ... is the, the labs, uh, you know, obviously have this immense amount of compute power and more advanced models than- 00:24:58,580 --> 00:24:58,900 [Eric] Mm-hmm 00:24:58,900 --> 00:25:16,820 [John] ... than the average person has access to. And, um, it comes down to currently a little bit of a, um, difference in capability of, of the de- in this case, the defender, like what tools do they have to defend, versus the attacker and what tools do they have to attack. 00:25:16,820 --> 00:25:16,900 [Eric] Yep. 00:25:16,900 --> 00:25:20,420 [John] Even though it wasn't obviously not intentional at all, and they worked together to solve it. 00:25:20,420 --> 00:25:22,540 [Eric] Sure, it wasn't malicious. Yeah. 00:25:22,540 --> 00:25:28,780 [John] Yeah. But that- that's one of the big debates, and I think that's what has some people scared right now. 00:25:28,780 --> 00:25:29,560 [Eric] Yep. 00:25:29,560 --> 00:25:36,760 [John] Is to the average company, my tools are not as good to defend myself as the tools, you know- 00:25:36,760 --> 00:25:36,880 [Eric] Yeah 00:25:36,880 --> 00:25:37,780 [John] ... in some, in some of the labs. Um. 00:25:37,780 --> 00:25:40,660 [Eric] Yeah, absolutely. And I think that's legitimate, right? 00:25:40,660 --> 00:25:40,780 [John] Yeah. 00:25:40,780 --> 00:25:43,040 [Eric] I think one of the c- conclusions here is that, 00:25:44,220 --> 00:25:46,260 [Eric] uh, it, AI 00:25:47,540 --> 00:25:49,200 [Eric] increases the, 00:25:50,820 --> 00:25:57,540 [Eric] uh, capacity and complexity of the attacks that can be run. 00:25:57,540 --> 00:25:58,340 [John] Right. 00:25:58,340 --> 00:25:59,690 [Eric] Uh, I, like, maliciously, right? 00:25:59,690 --> 00:25:59,700 [John] Right. 00:25:59,700 --> 00:26:06,939 [Eric] So you could, you can imagine someone saying, "I want to hack Hugging Face," you know, for whatever reason. 00:26:06,940 --> 00:26:07,620 [John] Right. 00:26:07,620 --> 00:26:09,200 [Eric] Or worse, you know, hack into 00:26:10,260 --> 00:26:13,280 [Eric] a, you know, fintech startup that has a bunch of people's, like- 00:26:13,280 --> 00:26:13,540 [John] Sure 00:26:13,540 --> 00:26:14,290 [Eric] ... bank account information. 00:26:14,290 --> 00:26:14,320 [John] Yeah. 00:26:14,320 --> 00:26:14,580 [Eric] Right? 00:26:14,580 --> 00:26:15,360 [John] Right. 00:26:15,360 --> 00:26:22,800 [Eric] Uh, and so the tools at their disposal with AI are, are far more powerful than any of the tools that they have had previously. 00:26:22,800 --> 00:26:23,320 [John] Right. 00:26:23,320 --> 00:26:36,830 [Eric] Um, that should be a cause for concern. I think there are two levels. Like, one, uh, if you work in tech and you have a software product, you absolutely should be scanning for, for vulnerabilities- 00:26:36,830 --> 00:26:36,830 [John] Right 00:26:36,830 --> 00:26:38,960 [Eric] ... and be on the offensive. 00:26:38,960 --> 00:26:39,480 [John] Right. 00:26:39,480 --> 00:26:39,520 [Eric] Um- 00:26:39,520 --> 00:26:46,930 [John] And, and then that... Yeah, and I think the big takeaway there is not... There, there is a practical takeaway, 00:26:47,940 --> 00:26:52,840 [John] and it's, it used to be okay to be up to the month, 00:26:54,140 --> 00:26:55,010 [John] maybe week- 00:26:55,010 --> 00:26:55,010 [Eric] Mm-hmm 00:26:55,010 --> 00:26:59,459 [John] ... for vulnerabilities, and I think instant patching is gonna be more important. 00:26:59,460 --> 00:27:00,320 [Eric] Yeah. Yep. 00:27:00,320 --> 00:27:03,619 [John] Because man- many companies would be years behind in patching vulnerabilities. 00:27:03,620 --> 00:27:04,360 [Eric] Totally. 00:27:04,360 --> 00:27:06,140 [John] So that's a really practical thing. 00:27:06,140 --> 00:27:13,940 [Eric] And there are actually tools. Like Vercel released an open source tool called DeepSec, uh, we can link to in the show notes, that you can just actually run. You can just- 00:27:13,940 --> 00:27:14,190 [John] Right 00:27:14,190 --> 00:27:17,700 [Eric] ... wire it up to AI gateway, pick a model depending on how much you wanna spend- 00:27:17,700 --> 00:27:17,800 [John] Right 00:27:17,800 --> 00:27:19,920 [Eric] ... and it will do, like, a very deep vulnerability- 00:27:19,920 --> 00:27:19,930 [John] Right 00:27:19,930 --> 00:27:24,140 [Eric] ... scan. So the good news is there are tools being developed that help you- 00:27:24,140 --> 00:27:24,420 [John] Right 00:27:24,420 --> 00:27:26,660 [Eric] ... sort of proactively defend yourself. Um, 00:27:28,480 --> 00:27:29,600 [Eric] but I think the other... 00:27:30,680 --> 00:27:39,850 [Eric] So yes, very serious. I think that we do need to take that seriously. The other side of it is that it, it happens on a personal level too, right? So if you- 00:27:39,850 --> 00:27:39,850 [John] Yeah 00:27:39,850 --> 00:27:57,080 [Eric] ... think about phishing, uh, if you think about, uh... I mean, someone in our family, um, you know, elderly, uh, their spouse had died, and, you know, they still, they still live alone, like, a decent bit of the time, 00:27:58,140 --> 00:28:00,600 [Eric] and they were subject to a really malicious, you know- 00:28:00,600 --> 00:28:00,809 [John] Mm-hmm 00:28:00,809 --> 00:28:02,780 [Eric] ... AI attack that happened through social media- 00:28:02,780 --> 00:28:03,120 [John] Yep 00:28:03,120 --> 00:28:13,920 [Eric] ... and ended up, you know, losing a bunch of money because of it. Um, and it was unbelievably difficult to tell, uh, that exactly what was happening, especially for that person. 00:28:13,920 --> 00:28:14,089 [John] Sure. 00:28:14,089 --> 00:28:22,600 [Eric] And so that is another thing I would say is, like, you know, pa- like, storing your passwords in an open Notes document. 00:28:22,600 --> 00:28:22,610 [John] Yeah. 00:28:22,610 --> 00:28:28,100 [Eric] You know, all those sorts of things are becoming much, much more risky now because AI is getting- 00:28:28,100 --> 00:28:28,149 [John] Yeah 00:28:28,149 --> 00:28:30,440 [Eric] ... better and better at finding those vulnerabilities- 00:28:30,440 --> 00:28:30,450 [John] Yeah 00:28:30,450 --> 00:28:31,800 [Eric] ... even just in your personal life. 00:28:31,800 --> 00:28:37,140 [John] Yeah, my friends and family advice around that is, one, get some kind of password manager- 00:28:37,140 --> 00:28:37,210 [Eric] Yeah 00:28:37,210 --> 00:28:37,610 [John] ... for sure. 00:28:37,610 --> 00:28:38,520 [Eric] 1Password. 00:28:38,520 --> 00:28:39,500 [John] Yeah, that's what I use. 00:28:39,500 --> 00:28:39,680 [Eric] Yep. 00:28:39,680 --> 00:28:57,000 [John] Yeah, 1Password. And, and the second one, when I first thought about this, it was super encounter... It was very intuitive, but I hadn't thought of it. If you, um, reuse a password between two services, um, and one of those services gets hacked, then 00:28:58,040 --> 00:29:00,410 [John] the, you know, dark web, all the hackers- 00:29:00,410 --> 00:29:00,440 [Eric] Mm-hmm 00:29:00,440 --> 00:29:04,280 [John] ... have access to that, and will go try it on a bunch of other services- 00:29:04,280 --> 00:29:04,290 [Eric] Yes 00:29:04,290 --> 00:29:08,439 [John] ... and often get lucky and can get into other services. 00:29:08,439 --> 00:29:08,780 [Eric] Yes. 00:29:08,780 --> 00:29:18,500 [John] I just don't... And, and password managers help you have unique passwords per thing, but that is probably one of the number one things that people don't do, is have a unique pass- 'cause it's annoying unless you have- 00:29:18,500 --> 00:29:18,509 [Eric] Mm-hmm 00:29:18,509 --> 00:29:23,300 [John] ... a password manager. To just have a unique password for every single separate thing that you use. 00:29:23,300 --> 00:29:23,980 [Eric] Yep. 00:29:23,980 --> 00:29:26,920 [John] So... And you don't always know something gets hacked, right? Like- 00:29:26,920 --> 00:29:27,399 [Eric] Mm-hmm 00:29:27,399 --> 00:29:31,310 [John] ... like, c- and companies don't even know half the time nowadays if they've been hacked or not. 00:29:32,330 --> 00:29:32,380 [Eric] Yep. 00:29:32,380 --> 00:29:37,220 [John] So that way, if your username and password gets out, that password's useless. It's never been used any other place. 00:29:37,220 --> 00:29:37,580 [Eric] Yep. 00:29:37,580 --> 00:29:49,660 [John] That's probably the number one thing. Um, and, and then of course, like almost every service nowadays has that multi-factor. It'll text you, it'll whatever, which is annoying to set up, but super important. Like, that's the second- 00:29:49,660 --> 00:29:49,670 [Eric] Yep 00:29:49,670 --> 00:29:51,680 [John] ... those are the two very basic things I would say. 00:29:51,680 --> 00:29:57,139 [Eric] My very basic advice is do not give your bank account information to a Nigerian prince. 00:29:57,139 --> 00:29:58,180 [John] [sighs] You know- 00:29:58,180 --> 00:29:58,630 [Eric] That's [laughs] 00:29:58,630 --> 00:29:59,540 [John] ... should've thought of that. 00:29:59,540 --> 00:30:13,400 [Eric] There's, if there's one thing [laughs]... Uh, okay. Well, cybersecurity is definitely heating up, uh, and both for tech companies, you know, anyone with software, anyone running a database, uh, anyone- 00:30:13,400 --> 00:30:13,570 [John] Yeah 00:30:13,570 --> 00:30:22,509 [Eric] ... who has sensitive information in systems, uh, definitely need to think about that, need to think about it on a personal level. But also, AGI is not here. 00:30:22,509 --> 00:30:22,519 [John] Right. 00:30:22,520 --> 00:30:23,880 [Eric] The behavior of what happened- 00:30:23,880 --> 00:30:23,890 [John] Yeah 00:30:23,890 --> 00:30:25,820 [Eric] ... is expected, and I think- 00:30:25,820 --> 00:30:25,890 [John] Yeah 00:30:25,890 --> 00:30:27,480 [Eric] ... it shows us how impressive the models are. 00:30:27,480 --> 00:30:28,079 [John] Yep. 00:30:28,080 --> 00:30:35,860 [Eric] But, uh, but it's not AGI. This is just a, you know, sort of standard, um, standard fare for things that go wrong in tech. 00:30:35,860 --> 00:30:36,920 [John] Yep. 00:30:36,920 --> 00:30:41,620 [Eric] All right. There you go. GPT-6 breaks out. Be careful. 00:30:41,620 --> 00:30:42,199 [John] Yeah. 00:30:42,200 --> 00:30:42,880 [Eric] And don't worry. 00:30:42,880 --> 00:30:43,230 [John] Stay safe out there. 00:30:43,230 --> 00:30:45,470 [Eric] Don't worry about AGI. [laughs] 00:30:45,470 --> 00:30:45,480 [John] [laughs] 00:30:45,480 --> 00:30:55,210 [Eric] We'll catch you on the next show. [outro music]