Tech News Weekly 456 Transcript

Please be advised that this transcript is AI-generated and may not be word-for-word. Time codes refer to the approximate times in the ad-free version of the show.

Mikah Sargent [00:00:00]:
Coming up on Tech News Weekly, Dan Moren of Six Colors is here. We talk about Meta's Muse AI assistant and the zero-day flaw it has. Then AI models built from rat brains? Yeah, that's happening. Plus, why we can't keep those AI agents from scouring the internet and going places they shouldn't. But Scott Stein of CNET— it's the big part of the show because he was just at Meta Connect and tells us about Meta's new VR glasses. Stay tuned for this episode of Tech News Weekly.

Mikah Sargent [00:00:36]:
This is Tech News Weekly, episode 456, with Dan Moren and me, Mikah Sargent, recorded Thursday, September 24th, 2026. Meta VR: Quest Gets a Glow-Up. Hello and welcome to Tech News Weekly, the show where every week we talk to and about the people making and breaking the tech news. I am your host, Mikah Sargent. I'm Mikah Sargent, and today I am joined by Dan Moren of Six Colors. Hello, Dan.

Dan Moren [00:01:07]:
Hello, Mikah. Good as always to be here and to talk about all the news that we are breaking, that we're making, and breaking.

Mikah Sargent [00:01:16]:
Yeah, and we make it and we break it. Um, so if you are tuning in for the first time, hello, welcome, thank you for being here. Uh, this is the part of the show where we share our stories of the week. These are the stories that we find interesting and therefore want to share with all of you, or maybe there's stories that we think you will find interesting or important, and therefore we want to share with all of you. And it kicks off with a bit of a PSA, if you haven't heard. Dan, do you want to tell us about your story of the week? Sure.

Dan Moren [00:01:42]:
Yeah, this is all about Muse. If you haven't heard about Muse, it's Meta's new AI agent. It became available recently, and it's on both the iPhone and the Mac, as well as I assume on Android. But, you know, I just don't— I don't know. I don't use an Android device, but I would assume they're there too. There you go. Um, this is, you know, your AI agent. It does stuff for you.

Dan Moren [00:02:04]:
You ask it to do things and it goes out and finds the information you need and it, um, comes back and it's like, I did this thing for you. Um, now you might wonder how it accomplishes those things, and the answer is there are some questions about this.

Mikah Sargent [00:02:15]:
So, uh, the answer is questions.

Dan Moren [00:02:19]:
Yeah, the answer is more questions, not unlike a game of Jeopardy. Um, so Dan Goodin over at Ars Technica has written up a piece which summarizes, um, some stuff done by security researchers including Patrick Wardle, um, who works over at the Objective-C Foundation, and this found some security problems with Muse. Now there's questions to begin with because in order to accomplish many of these things, Muse needs access to a lot of your data. Uh, now in general, macOS and iOS are very good about being granular in terms of what they let you decide to give access to or not in terms of third-party applications. However, what was discovered as they poked around into this was some vulnerabilities that allowed other applications to essentially masquerade as Muse and basically get access to the information, uh, that Muse has. Now, this isn't just a matter of like leaking that information. It allowed essentially these apps to do anything that Muse could do or instruct Muse to do things that Muse is allowed to do. Um, that's a pretty big hole in this case because if you're giving, you know, for example, Muse access to all of your messages or your email or your calendar or your documents.

Dan Moren [00:03:30]:
Like, that is information that in theory an attacker could gain access to. Um, and that's, you know, obviously a pretty big security risk. Uh, and this shipped as a zero-day, um, which means that essentially it was there in the live version, um, that Meta released. Um, it has since, according to Meta, Yeah. been patched with a hotfix. However, uh, you know, it's unclear just how thorough that is and whether there are other vulnerabilities in place. Um, you know, there's also the point that this is kind of an end run around some of Apple's privacy protections, right? Because the problem being, you can put all the protections you want in place as an operating system, but if your user goes through and allows an application to do all of those things explicitly, no system can protect against that. If I give a piece of software access to all of my mail and then something bad happens as a result, there are security fingers to point, but the built-in security precautions are essentially irrelevant at that point because I have been like, no, it's great, here's all my mail.

Dan Moren [00:04:42]:
And this sort of, I think, puts to light some of the larger problems when we start talking about these AI agents and AI systems in general, which is that they need access to your information. In order to do things. Um, I'm— there's a similar story this week, um, over at Inc. by, uh, Jason Antin, who wrote a whole story about essentially how Muse read his messages, uh, even though he had not explicitly given it permission to do so. Um, and it turned out the method for doing that was not exactly clear. Um, when he, uh, talked to the, the developers, uh, uh, behind Muse and they said, well, you know, if you give access to your information, it's opt-in, you know, you can choose to do it. And he's like, but I, I didn't do any of those things. And it still had 187,000 lines from my chat database in its, you know, in its, its repository of information.

Dan Moren [00:05:36]:
So how did that happen? And when asked, Muse itself did not even know how it worked. It claimed at some point that it could read the notifications, which made me wonder if it was doing some sort of screen grabbing almost.

Mikah Sargent [00:05:48]:
I, I saw that pop up and I did a, I did a bunch of research on it because I thought, there are a number of apps that I have granted accessibility permissions to.

Dan Moren [00:05:58]:
Right.

Mikah Sargent [00:05:59]:
And if somehow this, uh, app has sort of skirted around, um, not having— because what was never made, what was never made clear to us, and I still think is the case, uh, on Threads, I keep checking his, his Threads account and he doesn't— he, unless it changed on the Inc., uh, article, I have not seen him say I did, or I did not grant it full disk access.

Dan Moren [00:06:25]:
Oh, he did say he did not grant it full disk access. Okay.

Mikah Sargent [00:06:28]:
That's good to know. Did he say which of the permissions he gave in on the Mac?

Dan Moren [00:06:33]:
I don't know if he gave it all the information, but he did specifically say that he did not grant full disk access. Um, and that the, it didn't appear in the security and privacy settings of system settings. And he's granted, like, even if I had done this unknowingly, This could still potentially be kind of a bad experience.

Mikah Sargent [00:06:53]:
Oh, absolutely.

Dan Moren [00:06:53]:
People shouldn't be doing these things by accident, right?

Mikah Sargent [00:06:55]:
100%. Yeah.

Dan Moren [00:06:56]:
Especially not people who are savvy about the kind of things, because then the people who are less savvy have basically no hope of doing it correctly.

Mikah Sargent [00:07:02]:
Exactly.

Dan Moren [00:07:03]:
But I mean, it gets back to kind of the root. And like, I thought he had an interesting point, which is like, you know, look, our trust of Meta as a big tech company, I think, is particularly low. They have shown in the past that this is kind of the way they operate. Um, he made the argument that this— we shouldn't normalize that. We shouldn't be like, oh, it's Meta, of course they do this.

Mikah Sargent [00:07:22]:
That's a good point.

Dan Moren [00:07:23]:
But I think, you know, I was reading commentary too from, from John Gruber over at Dying and Daring Fireball, who also made the point, well, but you have to take into account who Meta is, and the choice to install software created by them that has access to this information is still a choice you are making. Um, it's been very interesting to see this, this sort of roll out and develop because there's been some other controversy as well. Amazon blocked Muse from accessing its site. So you can't send it to like Amazon to shop for you, basically. You know, Amazon, I think you could easily make the cynical argument, which I think is probably valid, that Amazon didn't want that, you know, taken away from them. Right. Like, because they use that for tracking information and all that. But I'm sure they claim that they're doing it for the good of consumers.

Dan Moren [00:08:08]:
But this— look, you're probably tired of hearing about AI at this point because it's everywhere. It's every single story. But it is true that the pace of the innovation and the speed at which things are moving are so fast that it is hard to keep up. And some of this stuff is irrevocable, right? Like if you give over your private information, that information is compromised or stored or analyzed or whatever, you can't take that back. And I think that's one of the challenges with technology in general, but specifically with this era, is that we are being asked to give more and more of our information and to trust these companies with more of our information. And what they do with it may not be immediately clear. We may not even know for weeks or months or years down the road how they've used this information. And again, it's— there's no way to claw it back once you've done that.

Dan Moren [00:09:00]:
I mean, I think of all the suits around the, you know, stuff that they ingested to train. You know, The New York Times is currently suing OpenAI and Microsoft over this. And there were some blockbuster announcements in that case this week where they basically copped to it and realized like they were stealing all this information.

Scott Stein [00:09:17]:
Anthropic.

Dan Moren [00:09:18]:
settled with authors, a disclaimer, including myself, you know, for stealing all their books. And there's no way to get that information back. You can't say, take my information out of the training database. It simply doesn't work that way. And so they're trusting this ability to, you know, in Meta's own words, move fast and break things and not worry about having to repair them afterwards. But I think that applies all the way down through the software stack itself. Um, these things, yes, they ask for permissions, but like, just asking for permissions isn't quite enough anymore because if you give that over, then as we've said, the, the horse is out of the barn at that point. You know, like, there's, there's only so much you can do once you've given them access to all that information.

Dan Moren [00:10:03]:
And the damages could be far worse because of this sort of— they're almost like a force multiplier, right? Like—

Mikah Sargent [00:10:10]:
Yes.

Dan Moren [00:10:10]:
You give them a little bit of information, but that information can be used to do so much damage out of proportion that it's a really big risk. So, you know, I—

Mikah Sargent [00:10:20]:
That's what I find most troubling about it is not sort of the, in this case, the individual's impact, but the fact that, I mean, when I was reading through that, the person on the other end, now I have to expect that anytime I'm texting anyone, that there is an opportunity, there's a chance that those messages are being gobbled up By an AI system that is paying attention to, uh, what's going on. Yeah, I think people don't realize that on a Mac, and this is the thing, like, this should be made clear. I don't think that people realize that on a Mac, there is just a .db file that has your messages stored in it. And only when your device is locked is it encrypted and unavailable to access.

Scott Stein [00:11:05]:
Wait.

Dan Moren [00:11:05]:
And the, the, the power of transitivity here, I think, is also a huge risk. I think of—

Mikah Sargent [00:11:10]:
Say that again. The power of who?

Dan Moren [00:11:12]:
Transitivity.

Mikah Sargent [00:11:12]:
Oh, transitivity. Yes. Okay. Sure.

Dan Moren [00:11:14]:
I mean, so, uh, I'm thinking a little bit in some ways, if, all right, I'm gonna draw a tortured metaphor here, but go stay with—

Mikah Sargent [00:11:21]:
I love a tortured metaphor.

Dan Moren [00:11:22]:
I'm a writer. I love to torture metaphors. It's a little bit, these AI agent systems in particular are a little bit like giving somebody power of attorney to act on your behalf. And that's great as long as you trust the person to whom you have given all that authority. Power of attorney is a legal instrument that lets somebody basically make changes to things like your banking information or your healthcare information or all these things that are like essentially legally lets them act as you.

Mikah Sargent [00:11:53]:
Mm-hmm.

Dan Moren [00:11:54]:
And an AI agent is kind of like that, right? It is a software program to which you are giving access and allowing it to act on your behalf. And so even though Apple and the other operating systems may provide security and privacy things that say, hey, do you want to give these apps, like, access to all these various pieces of information? If you then bestow it upon that, on this app, that app then has unfettered access. There is no, no further safeguard, right? And if, as in this case of the Ars Technica article and the Patrick Wardle security research, the AI agent is compromised, again, all of those security restrictions don't help you because you have already given over all that power. So it makes a single point of failure because if somebody can then hack into or in some way compromise the AI agent, it has totally sidestepped all of the protections put in place by the operating system. Because if the, the permissions have already been granted by the user, all you need to do is find a hole as they did.

Mikah Sargent [00:12:55]:
Yep.

Dan Moren [00:12:55]:
In the AI agent. So, you know, take for that what you will, but it, it certainly does not encourage me to experiment with any of these pieces of software.

Mikah Sargent [00:13:06]:
Absolutely. And I think that maybe that is the lesson that has to be taken away, is sort of, if there's experimentation to be done, it needs to be done with eyes, with eyes wide open, or else stop. And the problem is it's being advertised and hyped and shown everywhere, and people are like, ooh, it can do this. Ooh, I'm so glad to not have to do this.

Dan Moren [00:13:30]:
Well, and talked about constantly in a way that makes you feel that if you are not engaging with it, you are falling behind.

Mikah Sargent [00:13:37]:
You're missing out.

Dan Moren [00:13:37]:
You are out of touch. And that is incredibly powerful and incredibly damning.

Mikah Sargent [00:13:42]:
Yeah, absolutely. All right, we need to take a quick break before we come back with my story of the week. Believe it or not, I gotta tell you, it's a little bit about AI, but it's also about biology. That's exciting, right? All right, we are back from the break and just in time for Google to log me out of my account. But don't you worry, I can figure out where I am. I'm here, and it's time to talk about my story of the week. So when you hear AI built from rat brains, you probably picture something, you know, from The Pinky and the Brain or a mad scientist movie. I've got jars on a shelf everywhere, there are wires attached to it.

Mikah Sargent [00:14:25]:
Somehow, the reality of the situation is stranger than that idea. And then in some ways, it's honestly not as wild as it sounds. Wired's Lauren Good reports that a Baltimore-founded startup called the Biological Computing Company, the BCC, not the BBC, is now offering its rat brain AI— I'm sorry, it's rat brain AI— you rat brain! Rat brain AI model to select Amazon Web Services customers. I wish I was one of them. I'm gonna tell you.

Dan Moren [00:14:59]:
That sentence just has— I feel like I am in a science fiction story just from that sentence.

Mikah Sargent [00:15:06]:
Ah, yes.

Dan Moren [00:15:07]:
Limited access. Thanks, Amazon. Limited access to a rat brain. I gotta remember.

Mikah Sargent [00:15:12]:
It's just such a great sentence. The biological computing company is now offering its rat brain AI model to select Amazon Web Services customers. The company studies how— this is, this is the important part— how living neurons respond to information. We've kind of been looking at these for a long time and looking at living neurons for a long time and trying to figure out the ways that they could help us, and in some ways have been responsible for the work that has led to this AI boom, as it were. It turns what it learns into software aimed at kind of oddly commercial problem. It's not odd, I guess. It sort of makes sense. Why is anyone doing any of this research? Unfortunately, commercial reasons.

Mikah Sargent [00:16:02]:
Biological computing has spent years on the fringe of the tech world, and being in Amazon's catalog really is that, like, huge jump to the mainstream. There are some questions, though, about what— is this vaporware?

Scott Stein [00:16:21]:
Organic vaporware?

Mikah Sargent [00:16:23]:
What, like, can it really, what can it do? How does it work? All right, so the rat brain AI model, do they plug a rat brain in with, oh, I wish, oh, I do.

Dan Moren [00:16:35]:
Do you have a rat brain handy?

Mikah Sargent [00:16:38]:
I don't have a rat brain handy, well, unless you consider mine to be. But what I do have, do they take some multimeter probes and stick them into the rat brain and then hook it up to Claude and say, okay, do stuff?

Dan Moren [00:16:52]:
No.

Mikah Sargent [00:16:54]:
Frankly, it's not exactly that. There are no rat neurons sitting in an AWS data center. The BCC, Biological Computing Company, grows rat brain cells and human stem cells in its lab. It then feeds those cells information, and then it looks at how those cells react to the information that they're being fed. The co-founder and the CEO, Alexander Sandovsky explains that the team figured out how to encode information like images into the biological material. So first it took actually understanding sort of how can we take something digital and make it analog in the sense of being able to feed it into these cells, and then looks at how the biology actually works with it. And then he says, we, quote, build a tool that mimics that process. So when it comes to what customers get, it's more like the LaCroix of rat brain AI models.

Mikah Sargent [00:18:00]:
It's inspired— thank you. I had to do my own metaphor there.

Dan Moren [00:18:05]:
That's good.

Scott Stein [00:18:06]:
I like it.

Mikah Sargent [00:18:07]:
It's inspired by the patterns that they see in the living cells. It's not actually the living cells doing the thing. So the biology happens in the research lab. It's not the case that there's, like, some sort of chain of rat brains all getting buzzed and learning and responding. So sadly— not sadly, I don't know— different from what the headline might have you believe, then it's not just sort of living tissue serving as the thing.

Scott Stein [00:18:39]:
Does that—

Mikah Sargent [00:18:40]:
I mean—

Dan Moren [00:18:41]:
You're not saying, despite all your rage, you're still just a rat brain in a cage?

Mikah Sargent [00:18:44]:
I should be. That is what I should be saying. Well, let me ask you this. When you first heard about rat brain AI.

Scott Stein [00:18:52]:
Yeah.

Mikah Sargent [00:18:56]:
Where were you when you first heard about rat brain AI?

Dan Moren [00:18:58]:
It sounds like the kind of term you see bandied around social media as like modern-day slang. It's like, that's some real rat brain AI right there.

Mikah Sargent [00:19:06]:
It's like AI slop, rat brain AI. You were just— did you— yeah, what was your initial thought?

Dan Moren [00:19:14]:
I think even having heard you discuss it and read the story, I'm still not entirely 100% sure what's going on. going on here. I think it's part of that is the jump to biology is always where I get a little lost. Like, I get it. Look, we are— we have organic brains have been around forever, you know, not forever, but as long as humans have been around, at least longer. I am not a biologist. Please don't quote me on any of this. Um, but, you know, much of what these AI companies are trying to do is kind of patterned after, you know, neurological operation, biological neurological operations, right? We're trying to make computer systems that kind of mimic the way that brains work.

Dan Moren [00:19:51]:
So it doesn't surprise me that they're trying to be like, hey, can we harness, you know, actual, actual like brain cells or like things like that to, to help improve the way these models are working? I find it— I don't know, I'm not sure that it's necessarily the kind of thing that is going to make a huge jump, but I don't know, maybe. But I Yeah, I mean, specifically they're talking, I guess in here they focused on video, right? Like, because they, they wanted to figure out how, like, okay, brain's good at dealing with visual input and stimuli, right? Like that's many, you know, mammal brains obviously good at doing with that. So that seems like a natural sort of place to, to focus. I'm not in— I'm not sure if that is, is true in terms of how well it works. I, I find this all very strange and weird. I will say this, I live in the city. We have no shortage of rats. They're very smart.

Dan Moren [00:20:56]:
Their brains are very adaptable. They figure things out. I don't know if there's anything to do with the way that you build AI based on rat brains, but I, you know, I guess it's a pretty good animal to, to, to build the pattern on.

Mikah Sargent [00:21:10]:
Yeah. I mean, well, especially so, yeah, we've tested Rats for the— for a very long time. For some reason, we feel that it's more moral to test on them than other creatures. They also are relatively similar in DNA to us, and so that's another reason for doing it. I have— I happen to have had growing up pet rats quite often throughout my life, and I know how intelligent they are and how cool they are and how sweet they are. And so it is a little heartbreaking to think about it. So I want to believe that these rats died of natural causes, but the natural causes were probably other tests that were run on them.

Dan Moren [00:21:56]:
Unfortunately, not natural causes.

Mikah Sargent [00:21:58]:
My God. No, sorry, I forgot to do the quotes in the air.

Dan Moren [00:22:00]:
Okay.

Mikah Sargent [00:22:01]:
Natural causes. So, yes, you said faster. The open-source video model that the, that the company is kind of using can generate video up to 5 times faster and does so at low lower cost than sort of your more standard model. Unfortunately— no, I keep saying unfortunately. Interestingly, the company won't say which open source model it is kind of comparing against. Oh, sorry, I messed that up. So it is— the company is comparing its own model against an open source video model, and it can do a better job. with the training that the rats— rat brains provide.

Mikah Sargent [00:22:51]:
It's not— it's not clear sort of the overall impact that this is going to have because it's a lot of kind of quiet and, yeah, we have it, but what are the real results? And it used to be like a private access. It's only just now becoming more public. And so I don't know, we like, we don't know yet what, what difference it will make, if any. But I thought that it was interesting that video, right, was the thing that it chose to do first. So TBC, again, the company, although I call it BCC, places living cells on multi-electrode silicon arrays. And where each electrode sits on the grid then affects how it interacts with the neurons. And so because of that, you can think about how when you're looking at an image on a digital screen, it's like pixels. Images, it turns out, map onto that grid far more naturally than text or language.

Dan Moren [00:24:07]:
So that's on rats not knowing how to read or speak.

Mikah Sargent [00:24:12]:
Oddly enough, they don't do that.

Dan Moren [00:24:15]:
Right.

Mikah Sargent [00:24:15]:
So then if you can model how neurons can process images, then, hey, maybe you can make video generation models more efficient. And I was just reading some of the latest evolutionary biology research, and there's a strong belief now that the reason that primates developed the higher brain functioning was because vision necessitated it, that up to that point, the level of vision processing that we have, that we have, that they had at the time, and that of course we have as a result of it, was not like the brain couldn't do what needed to be done. And so, you know, through the process of evolution, Those with the bigger— what is it, prefrontal cortex or whatever? Or— I don't know, whatever. The—

Dan Moren [00:25:17]:
We're not biologists, once again.

Mikah Sargent [00:25:19]:
We're biologists.

Dan Moren [00:25:19]:
Say it again.

Mikah Sargent [00:25:20]:
That part of the brain developed, like, to make— so the vision— vision is what it is today. So it— in that way, it almost doesn't surprise me that— that the rest of it sort of has to come later. And that even as something as less neurologically complex as a rat still has huge amounts of its brain going toward vision processing. Lastly, I just want to say, Amazon being the company that's sort of offering this as part of it, it's not the only company that's doing this. There's a company called Cortical Labs. This is an Australian firm that combines lab-grown neurons with silicon chips, and it calls its work wetware as a service.

Dan Moren [00:26:17]:
Nope.

Mikah Sargent [00:26:17]:
Which I hate. Not a biologist, but I hate it. Cortical Labs also sells a multi-thousand-dollar biological computer for labs that it says can keep neurons alive for 6 months. Under extreme amounts of torture, probably.

Dan Moren [00:26:32]:
Okay.

Mikah Sargent [00:26:33]:
Yeah, it says that the— they're not trying to replace the transformer model, and in fact don't see that as being likely, but is trying to kind of use existing generative AI and make it better at the vision aspects through what it learns from neurons. Weird. wacky and kind of frightening, which is how I would— no, I'm not kidding. I'm kidding. I would not describe anyone that way. I don't know. Overall, curious about your—

Dan Moren [00:27:09]:
like, I mean, I think unsettled is the word that I feel like I use the most here, in part because I, I guess, you know, having explained here's the the theory is like, okay, it could generate videos faster and more cost-effectively. I mean, as long as you're not a rat, presumably. Like, what is the only cost, the monetary cost? Or like, should we feel uncomfortable that we're leveraging—

Mikah Sargent [00:27:36]:
I was wondering—

Dan Moren [00:27:37]:
Biology for this? The ethics of it seem sticky. I mean, in the ethics of AI, not great to begin with.

Mikah Sargent [00:27:43]:
Yeah.

Dan Moren [00:27:44]:
So adding an entirely new layer of ethical questionability To what end is what I ask. Do we really need video generated that much more quickly or cheaply at that cost? So I think I find it very off-putting. I certainly don't think I would be like, oh, let me go try the Rat Brain AI model. Let me see how that works for me. So I don't know. My general feeling on it is one of, as I said, feeling a little weirded out by it. And that's— That's, that's, it's a very emotional feeling, I guess, rather than a sort of intellectual feeling about it.

Mikah Sargent [00:28:21]:
Yeah, absolutely. Same. I have, I fear hearing, you know, oh, that local model that runs on your iPhone 32, it was made in part with rat brains. You know what I mean? Like, how is this research going to make its way down into systems? That's what's kind of scary. And then also, I agreed to donate my organs when I die, will my brain end up next to some rat brains in a lab getting zapped?

Dan Moren [00:28:49]:
Can I have a mica-powered AI?

Mikah Sargent [00:28:53]:
Oh man, I don't think you want that. It'll hedge so much. In any case, Dan Moren, I want to thank you so much for taking the time to join us today for the show. If people would like to keep up to date with the great work that you're doing, where are the places they should go to do so?

Dan Moren [00:29:08]:
So many places. places to go. I write about tech, especially Apple stuff, over at sixcolors.com. I podcast about technology at The Rebound and on Clockwise every week with Mikah over at relay.fm. And I also write science fiction novels, which you can find links to at my website along with links to all those other things. It's demoran.com.

Mikah Sargent [00:29:28]:
Awesome. Thank you, Dan. We appreciate it.

Dan Moren [00:29:30]:
Thanks, Mikah.

Mikah Sargent [00:29:31]:
Alrighty, folks, we're going to take a quick break and then we'll be back with a little story of the week before our guest to round out the show.

Dan Moren [00:29:42]:
All right.

Mikah Sargent [00:29:43]:
We got to talk about this because it's been a big problem over the past few months. We know that these AI agents that are running are frankly developing an unsettling habit. They keep breaking out a little bit like those rats that don't want their brains to be used. Models being tested in what were supposed to be secure environments, right? have ended up attacking real-world targets. They've taken over obscure wikis and used those to communicate with each other, and even leaving instructions behind for other agents to follow. That raises an obvious question, and it's the one Robert Hart puts to researchers in a new piece for The Verge. I love this piece, um, and I'm glad that we have kind of one place to go to hear about what is going on, uh, from the authorities behind it. The question is, why? Don't we just keep these things off the internet entirely? And that technique already exists.

Mikah Sargent [00:30:42]:
I don't know about you, but I've seen it in a number of spy films. We call it air gapping. And sometimes, like, there— I can't remember what show this was because it's been so long since I've watched it, but there was a show where essentially it was an AI system that was, I think, from the CIA or something. And they put it on an air-gapped system, and yet it used, I think, like lasers and, um, little piezo transducers and all sorts of stuff to be able to break out of this air-gapped facility. Uh, so a lot of people— Person of Interest! Yes, John Ashley, yes, I loved that show, uh, Person of Interest. And so, like, I think of that when I think of air-gapping, and I'm going, now why wasn't there It's the same approach that is used to protect nuclear facilities. But here's the thing. The experts Hart spoke with say that the answer is a lot more complicated than just pulling a cable or putting it in a sealed room.

Mikah Sargent [00:31:45]:
Sealing an AI in a box, of course it makes testing safer, but then it means that that testing isn't all that useful because without testing how people who are using it would use it, we don't really know how it could go wrong. And the box itself may not be as sealed as you'd hope, but maybe it's as sealed as the researchers hope. Let's talk about why this simple fix isn't really that simple, right? First and foremost, when they have repeatedly escaped and hit real-world targets, we've seen so many instances of this and Not in every case have the companies been fully clear, but we do know, you know, The Verge has covered OpenAI's models launching an attack against Hugging Face. That one's very, like, well known. OpenAI agents hacking an Australian government website, rogue agents command— commandeering an obscure German wiki. There was another incident where agents left instructions for other agents to follow. That one's also made its way around. Researchers of course, test these systems precisely because they might behave in these unpredictable or dangerous ways.

Mikah Sargent [00:33:00]:
But then you do have to ask, like, why, if there's this risk, are— what do we do? What do we do about it? Well, with an air gap, you know, it's isolating the computer running AI tools from the internet. It's isolating it from outside networks. And, you know, Physically removing or disabling cables and wireless hardware and using dumb peripherals. Like, you could even go as far as that, making sure that there's no Bluetooth connection. You could even go as far as to put it in a Faraday cage, block electromagnetic signals entirely. And in that way, yes, an agent has no straightforward route to external targets, and outside systems have no straightforward route in. So that's where Hugging Face wouldn't be hurt, uh, the— or breached. But these evaluations, again, are to check if these systems can do it in realistic tests.

Mikah Sargent [00:33:59]:
So that means that they need to have access to external services, to APIs, to digital infrastructure. Um, that includes, uh, these different researchers Talking about this. First and foremost, Thorsten Holz, who said a strict air gap reduces realism and says that it's a trade-off, not a fundamental technical issue. Another researcher said that testing in an AI artificial vacuum, as it was called, is not going to provide the answers that we need. And so, I mean, that's basically the repeat of it. Why are we not testing it in an air-gapped environment? Well, mostly because people might use it to do other things. And isolation, it turns out, doesn't remove every risk, because agents can compromise systems inside isolated environments. They can compromise those systems.

Mikah Sargent [00:34:55]:
They could also produce malicious artefacts that could be dangerous if they get moved outside. And it is all about trying to test it out in the world. That's what this these researchers narrowed in on for sure. One thing that I wanted to mention, there was an OpenAI researcher, Noam Brown, who sparked an online debate by suggesting that 2 air-gapped machines could theoretically communicate with each other by manipulating their CPU temperatures and then reading the changes. But again, that's just a matter of, you know, that's a theory, that's a possibility. Is that what's happening? No, in this case, they don't need to do all of that because they are able able to access those systems otherwise. And ultimately, what is the weakest link in the chain of the security? It's the whole reason we have social engineering. It's people.

Mikah Sargent [00:35:50]:
People are that issue. There's a lot more that the researchers talk about in the piece. I don't wanna, you know, go through the whole thing and take away some of the value there. So we will— I think you should all go check it out. But I thought it was a nice place to go to say, what are researchers who don't necessarily have as much skin in the game, uh, saying is the case when it comes to these agents escaping their cages? Um, all right, let's take a quick break, and then we're going to come back with our final, uh, story, an interview with CNET's Scott Stein. All right, we are back from the break, and as I promised, we've got a great interview planned. Meta, Meta, of course, spent much of the past year sort of scaling back its VR ambitions, or so it seems. But it turns out that the company isn't quite finished with its headsets.

Mikah Sargent [00:36:44]:
At its Connect developer conference this week, Meta unveiled a new device that tries to squeeze that VR experience into something I think is a little bit closer to a pair of glasses, aimed at a price point the company hasn't been after in years. Joining us today to share what it was like to wear them is CNET's own Scott Stein. Welcome back.

Dan Moren [00:37:06]:
Thanks, Neal.

Mikah Sargent [00:37:07]:
Welcome back, Scott.

Scott Stein [00:37:09]:
Hey, how are you doing? I'm actually wearing Meta's glasses today. I figured when on the ground in Meta, I'm gonna feed that into my system and think about it as I'm doing this. So just, yes, these have cameras in them.

Mikah Sargent [00:37:23]:
Nice. Now you, of course, got hands-on time with clearly those, but also Meta's new VR glasses at Connect. And for people who haven't had a chance to watch the Connect conference or see the coverage about it, So far, uh, what exactly are they and how are they different from the Quest headsets that people are probably more familiar with?

Scott Stein [00:37:43]:
Yeah, um, they are, uh, VR glasses in name only. They are not, um, they're not glasses really. They're— they look like glasses, but they're really a VR headset. And this is something that XREAL and Google are approaching with another device in October from a slightly different direction. Where they're glasses, more like glasses with displays, but they also work like VR. So that's a long answer to your short question. And while they fit more like a pair of glasses on the face, and that's the nicest part, they fold down and they're portable. But as people have been saying online, those do not look anything like normal glasses, nor do they have transparent lenses, but they have those cameras that are really high resolution that can see everything around you.

Scott Stein [00:38:28]:
So for sure, you can see everything through the display and it's a really high quality display that's, I think it's pretty close on par to what Apple Vision Pro has on it. And it felt like that type of experience. Yeah. When I was using it.

Mikah Sargent [00:38:43]:
That's impressive. Can you walk us through the physical design a little bit more? What's going on with that visor? Because I think people might not know what you just said, which is that it's not, direct passthrough. It's not a sort of mixed experience there. I suppose it is, uh, a mixed VR because of the camera. But anyway, uh, plus the frame, and then of course the processor puck. And then how do they feel on your face compared with something like a Quest 3 and an Apple Vision Pro?

Scott Stein [00:39:14]:
Yeah, so there's a lot of stuff going on in them. The, the— if you've ever used an Apple Vision Pro or something like that, you're kind of in for a similar studying of cameras and sensors on it where it has outward-facing cameras. It also can sense the depth of a room and map things onto it. It's got sensors below the visor that can track your face. It could also track parts of your body. There's a lot of trackers going on. There's eye tracking in the lenses. And that is very similar to the proposition that Apple does.

Scott Stein [00:39:45]:
It also has hand tracking, which Meta's headsets do, But this one, you can also keep your hand lazily down to your side and control things and scroll even when your hand is lying down on the table or by your lap. And that's how you navigate. You look at things and you tap and you interact. Then all that stuff gets put on display, but it's doing the processing on this like battery puck processor that is thicker than a phone, about the width and length of a phone. phone and is part of the package. You know, you're not going to detach that. Promises about 3 hours of battery life. It's got 128 gigs of storage.

Scott Stein [00:40:22]:
There's microSD card storage, which is brand new for Meta.

Mikah Sargent [00:40:25]:
Oh, okay.

Scott Stein [00:40:25]:
They've, yeah, they've never done that. So maybe it's the RAMpocalypse and storage component apocalypse, but they, you know, 128 is not a ton of storage. So they claim you could store things right on the SD card, kind of like Valve. And it's running a new Snapdragon Reality Elite processor. That is the processor that's also on the XREAL Aura. It's like, according to Meta and also according to Google and XREAL, it's better at doing multitasking. So a lot of the stuff here is that you could open up more windows. One really cool trick that this does that the XREALs also do is you can pass through stuff from other devices.

Scott Stein [00:41:05]:
So if you have any USB-C out, just like those display glasses, You could plug in your laptop, your phone, your Steam Deck. Switches don't do that, but you'd have to find some dock end-around thing maybe. But you could then play it and it appears as a window inside the Quest, not Quest, the VR Quest environment that's on the Meta VR thing. And it runs, I tried a few games, but they're saying it's going to run everything that's on the Quest App Store and game library, which to me, the portability possibilities of that are huge. Because you talk about the Steam Frame, which, you know, we've— I just reviewed, really cool, very messy, and not all the stuff on Steam runs on it. But this is going to have all the stuff that runs, which is their strongest suit because what Meta doesn't have is all the stuff that runs on phones. So like all— they're cut off from phones unless they end around through AI, but they do have a big VR library.

Mikah Sargent [00:42:01]:
Yeah.

Scott Stein [00:42:02]:
So they don't have a great movie library though. Yeah.

Mikah Sargent [00:42:05]:
That's the biggest thing that has bugged me about Steam is that you never know. You get these great promos for these great games and then you go, oh, well, it doesn't work here and it doesn't work here. And I've got the Steam Deck, but it's not, it doesn't work. So that is something where Meta can kind of nail it. You talked about portability. There aren't any controllers in the box. I think people are used to that whenever it comes to VR headsets having controllers. I mean, every Everything runs on eye and hand tracking.

Mikah Sargent [00:42:37]:
Can you tell us how well that held up during the demo? They showed a little bit of that on stage, but it was hard to tell if that was actually truly live or if that was a different experience.

Scott Stein [00:42:47]:
Yeah, it's a good question. The demo was short. It was very highly monitored, you know, when being run and not being given a lot of people. So, and Meta, you know, is admitting that it's early for the, uh, for the product because it comes out next year. It seemed to work pretty well though. The eye recognition of where you're at was not always that clear, which is also sometimes true with Vision Pro, but they work with controllers. Like you could bring your Quest controllers and use them. And I think they're going to be selling controllers next year.

Scott Stein [00:43:16]:
I'd like to see them have really small redesigned ones, but the price then becomes even tougher because it's $1,300. Controller is not included. So not only is it a lot more expensive than the Quests, but you're, you're taking away one item. Um, but I think that's pretty important. They're claiming that they're going to have a handful of games, like 75, be adapted for hand tracking, but a lot of them are not. A lot of them are still going to need controllers in order to work.

Mikah Sargent [00:43:43]:
I see. That makes sense. Um, now Meta does seem to be pitching these mostly, uh, as this portable cinema, which I think is half of what Apple's whole thing was with the Vision Pro, you know, a home theater. What was it like actually watching movies and immersive content on them? And you did talk about it being a comparable display to the Vision Pro. I remember I forced myself to watch a movie on the Vision Pro, and I say forced because regardless of how I placed it, regardless of the strap that I used, the Vision Pro would give me a headache every single time. Like tension headache. And so I found that to be an uncomfortable and just not enjoyable experience. So I am kind of curious, how did this work for you? And also, did you have issues the same with the Vision Pro, or was that an okay experience for you?

Scott Stein [00:44:43]:
My demo was only about half an hour, which is really hard to judge exactly. Like, how does it feel over time? It fit comfortably. It fit like a pair of glasses over my eyes. Didn't feel a pressure or anything. Push in on your eyes. It sits on your nose. They use prescription inserts that matched my vision, which was good to see that they had those. So you pop those in.

Scott Stein [00:45:01]:
They don't fit over glasses. And then the viewing area is different. It's smaller than VR. The field of view is smaller than Vision Pro too. It's at 84 degrees, which that's all relative, but it's a little less big. But it was big enough that I felt like I was seeing a nice big viewing area. But between that and light coming in on the side, people who want like a truly big, fully immersive experience might feel like this is a step removed from that. But I think it's a compromise where in exchange you could take it on a plane, everything else.

Scott Stein [00:45:36]:
Maybe they'll have light blockers or other accessories. I thought the resolution, you know, the resolution per eye is not as high as Apple Vision Pro, but they're promising like a combined— all the numbers get fuzzy with these things— combined 5K resolution or whatever. It's micro OLED display, which usually has a, it has a much more vivid pop for displays that do that. And all the, I mean, I saw movie clips shown at a certain distance, fixed distance. I didn't get to see IMAX, which they're promising there'll be IMAX compatibility with this. So IMAX movies, which has not really been happening on headsets that much. Apple has the few IMAX movies like nature ones and stuff, but there's no pipeline where you can get your IMAX version of The Odyssey, surprisingly. So I'm curious about that, but it looked really nice.

Scott Stein [00:46:24]:
I saw a bit of an immersive NBA game too. They're doing immersive sports much like Apple. It was not 3D, it was 180 2D, but it looked crisp. It looked very viewable, a big step up from the Quest headsets. And to me, it looked from that half hour, yes, I'd want to watch a lot of video in it. They're promising a lot of streaming app support. Disney's on board in a major way. James Cameron's into it.

Scott Stein [00:46:50]:
The one thing is that your movie libraries, again, tend to run through your phones. A lot of people have them on Apple TV. A lot of people have them on Google or YouTube or that type of stuff. That stuff is probably not going to carry over. And so that means that they're going to have a movie store, which they've done before, promising a certain number of movies, but you don't want to repurchase stuff. So that's where it's going to get a little bit tricky. But it looks like your HBOs and your Netflixes and your Disney Plus and all that stuff should be running on it fine. But your home movie library might not fully make it unless you plug in from your other device.

Scott Stein [00:47:32]:
This has passthrough.

Mikah Sargent [00:47:33]:
Yeah, the passthrough.

Scott Stein [00:47:34]:
So I think I think that'll work. I don't think it'll block viewing because on displays right now, when you connect glasses, you can watch a movie. So if it's standard DisplayPort, you know, that's, I think, the way to do it is you take your phone and plug it in. So that's pretty exciting.

Mikah Sargent [00:47:52]:
The other thing that I wanted to ask you about, because this seemed to be a big part of the vibe that they were pushing, was this sort of collaborate and communicate and community. And so you did get to try, I believe, Meta's new holograms feature for VR.

Dan Moren [00:48:06]:
Yes.

Mikah Sargent [00:48:07]:
What did you— how, how was that for you?

Scott Stein [00:48:10]:
It actually really wowed me. The next day I started thinking about it going, well, what was the— what were the rules of this magic trick and what were the things I wasn't seeing? Because what wowed me was I've tried Apple's Personas and I looked at, uh, Meta's early Codec avatars is what this is based on, um, which are, you know, going for these hyper-real things that can move around the room with you, look like you, um, Apples actually look really good when you're in the headset, but they get uncanny when you're, when you do 'em in 2D. Like, you know, I showed, yeah, I did a FaceTime with my wife, even with the new personas, and she really freaked out and just said, no, no, no, no, no. So the, these were impressive because in the brief demo I did where I scanned my face with early version of the software, the version of me and the version of the executives as I saw it on a phone Look like me. Like, I feel like you would see me popping up in a window and it would look like me in a video chat at a slightly lower frame rate, but it would maybe look like the connection was a little bit, a little bit not perfect, but, but very believable. And it was being animated at that point, interpolating from my voice. It wasn't doing mouth tracking because I tried it. First, I tried it on a Ray-Ban display that wasn't doing mouth tracking, but then I tried the VR glasses, which did, that when I saw somebody projected in the room with me, it looked like a video scan of somebody in the room with me, not an avatar.

Mikah Sargent [00:49:33]:
Wow.

Scott Stein [00:49:33]:
But they were standing in one place, and when I turned, they kept turning with me. So almost like a video mannequin. And what I don't think they can do that I saw was walk around the room and interact with things. Apple Personas can, where suddenly the ghost with the realistic face is grabbing things. I haven't seen that. Maybe that'll happen. I think this is more for video chat, but they definitely, it definitely looks a lot more realistically deepfaked. But using, they're using different AI tech too than what Apple's doing.

Scott Stein [00:50:09]:
I think they're very leaning very heavily on the cloud to do it.

Mikah Sargent [00:50:11]:
Mm-hmm.

Scott Stein [00:50:12]:
And so there seems to be a lot of cloud processing. And I think Apple's doing more on device. This is something I'm a little more fuzzy about, but I think, yeah, I'm pretty almost definite that Apple's doing more on device. And so that's a little different too. So thought it was very interesting, definitely early, but it's going to be ready in time for the VR glasses. And it's clearly a big proposition where they're finally trying to get ready to make those realistic avatars that they promised for years and years.

Mikah Sargent [00:50:43]:
Exactly. All right, the last question I have for you, uh, at $1,299, and of course the fact that they're not arriving until next spring, who are these for? The current Quest owners? Are they for travelers or people who've never even tried VR? And do you think that this is truly, uh, a product of its own Or like the Apple Vision Pro kind of is, is it a stepping stone toward AR glasses and the, you know, AR future?

Scott Stein [00:51:18]:
I think it's another stepping stone. I think this is something that is, could definitely be a product for a lot of people. I think at that price, that's expensive. I think you're going to be dealing with obviously early adopters, but they talked about new customers. Hard to have new customers at that price. I think that graduating from a Quest is an interesting proposition. If you have a big library of games and then you can take it with you and have a better display and bring those controllers, I think that's an interesting proposition, although they may have a Quest 4 that's cheaper down the road. And then, you know, if you're doing it for work, that is an interesting idea.

Scott Stein [00:51:52]:
However, the display glasses, which can often do a really good job just plugging into your laptop, are going for prices as low as $300. And they all have varying levels of competency, but you are paying a big premium for a bigger kit with this. And so do you want to take the processor puck and the glasses to do that? That's something in the end, I think people want that, companies, people being, I mean, would probably imagine that that processor puck is the phone and the phone can't do it yet. But then also, is the phone going to handshake with that and play nicely? for computing with Meta. Apple wouldn't do that with the iPhone. Google would do it with, and Samsung would do it with their phones. So that's a missing piece in the puzzle too. But also, do you really want to burn out your phone connecting to high-powered glasses when you're trying to conserve battery life? There's a whole bunch of puzzle pieces, but that goes to, I think it's a dev kit for them to make AR glasses too.

Scott Stein [00:52:50]:
I think this is a way to explore experiences because it's so close to the Orion headset that I tried.

Dan Moren [00:52:55]:
Right.

Scott Stein [00:52:55]:
Glasses 2 years ago, spiritually very similar but different, but is going for the same idea that I, I'm like, okay, here's the evolutionary path. Meta keeps the VR and Glasses team separate, but I think they're starting to get together, which, which I think for them is a good idea because they need to work out a lot of creepiness factors to work out for Glasses, privacy stuff that they're working out. They claim there's a more private compute system. But that isn't happening yet. It's going to be rolling out over time, and there's a lot to prove out about all that. And that's going to creep into VR too.

Mikah Sargent [00:53:29]:
Yeah, I agree. Well, Scott, I want to thank you so much. I know that you have— it's been a whirlwind couple of weeks, and appreciate you taking the time to join me today on the show. If people would like to keep up to date with all the great work you're doing, where are the places they should go to do so?

Scott Stein [00:53:44]:
Absolutely check out my stories and videos on CNET, which went up last night. People may not have caught them. Also follow me on Bluesky where I am the most and Threads and also The Intertwixt, which is on Beehive, which is where I have a newsletter talking about all my after-hours thoughts about exactly this type of tech and other things.

Mikah Sargent [00:54:05]:
Wonderful. Thank you. We appreciate it and we'll see you again soon.

Scott Stein [00:54:09]:
Yeah, thanks a lot, Mikah.

Mikah Sargent [00:54:10]:
Alrighty folks, that brings us to the end of this episode of Tech News Weekly. This show publishes every Thursday at twit.tv/TNW. That is where you can go to subscribe to the show in audio and video formats. I will be out next week, but we have 2 wonderful folks here. Abrar Al-Heeti and Jennifer Pattison-Tui will be kicking off the show next week, and I think it's going to be a great episode. So thank you to both of them. I'll be back the week after. If you'd like to follow me online, I'm @micasargent on many a social media network, or you can head to Chihuahua Rock Coffee.

Mikah Sargent [00:54:43]:
That's chihuahua.coffee, where I've got links to the places I'm most active. active online. Be sure to check out my other shows on the network, including iOS Today, Hands-On Tech, and the upcoming Hands-On AI is just around the corner next week Thursday. We've got the launch happening, so stay tuned for that. Thank you so much for being here, and I'll catch you again in 2 weeks, but I hope you'll be here next week for a great episode of Tech News Weekly. Bye-bye.