NC #1112 Eclipse by Steve, Allister's Vibe Coding Experience, Eddie on Clicks, Breaths, and Plosives, Security Bits with Bart Busschots

Hi, this is Allison Sheridan of the NoCillaCast podcast, hosted at Podfeet.com, a technology geek podcast with an ever-so-slight Apple bias. Today is Sunday, August 30th, 2026, and this is show number 1,112. Well, as you probably know, Apple has scheduled their next announcement event for Wednesday, September 9th at 10 a.m. Pacific time. That's a Wednesday this year because it's the week of Labor Day and we always get Mondays off in the U.S. for that, so it shifted it out to Wednesday. I bring all of this up to invite you to join other NoCillaCastaways to text chat about the event live while it's happening. The most important thing to know is that, as always, Steve and I will not be talking in voice during the announcement. We'll be chatting with everybody else in text. In order to join us, you can join our Discord channel at Podfeet.com slash chat at 10 a.m. on the 9th of September. As you also might know, Steve and I traveled to Spain recently to see a total eclipse of the sun. This was our fifth total eclipse, plus we saw one annular eclipse. Witnessing totality is something absolutely magical, indescribable, really, and this one did not disappoint. In fact, I don't know about Steve, but I think this was my favorite one so far. I actually cried a little bit, just a little tear in the eye. It was so emotional. Now, Steve loves to record and edit videos of eclipses, and he outdid it this time. The techniques that he used to create this are so interesting, I'm going to subject him to an interview about how he captured the eclipse with four cameras, a bunch of mics, stitching it all together at a later date. It's really, really impressive. But the important thing about his final video, as Bart says, is that he told a story with a two-minute video. It's not just the eclipse, it's how he spliced in the audience's reactions that make this such a compelling video. And again, it's only two minutes long. I put a link in the show notes to his video on YouTube, and I highly recommend you go spend those whole two minutes experiencing it. And I hope you're experiencing that eclipse with us. Bart wrote an extensive story about how to build a JavaScript command-line interface with Node.js, and he called this Tidbit 19 of Programming by Stealth. Now, it's a little confusing because tidbits aren't necessarily small, and in fact, this one is so big, it's actually going to be a three-parter. In Part B, recorded yesterday and published to the podcast feed, Bart walks us through the architecture of his Linkifier module. This is the linkifier module. This is how he creates all of the links you see in Security Bits and Let's Talk Apple. He starts with the three data modeling classes that he's been using for decades, but he explains how hard it is today to extract article headlines, which is the end goal, to go with the links. He walks through how he developed his extraction logic, but then had to do it per domain. It sounds crazy, but he uses surprisingly few sites for his articles for Security Bits and Let's Talk Apple. You can find Programming by Stealth Tidbit 19b in your podcatcher of choice, or you can follow him on Twitter. Follow the link in the show notes to listen along as you read it at pbs.bartofisser.net. Again, we've recorded Parts A and B, but not Part C, and he's noted it in the show notes exactly where we stopped, so you can keep up. All right, it's time to hand the microphone over to the dulcet tones of Alistair Jenks. On Episode 1009, Alison spoke about her addiction to AI vibe coding. She mentioned, many of her friends were also doing the same. I am one of those friends, and this is the story of my experience. First, a little background. I have been a programmer for around 45 years. I've done it for fun that entire time, and around six years after I started at age 12, my skills were good enough that it got me my first job. That job was the only time I have had programmer in my job title, but I have been coding as a working tool for a long time. In my personal life, I've made various hobby projects over many years, from building a photo database with its own sector-level floppy disk format, to hand-coding my first website in HTML, then converting it to use XML and XSLT, and eventually writing my first iPhone app. In short, I know how to code. My first iPhone app was built in my Christmas break at the end of 2014. It was very, very simple, but I felt a great sense of accomplishment. I'd had to learn Objective-C, but that wasn't the hardest part. No, the hardest part was the API. Actually, the hardest part was the App Store submission process, but that doesn't serve this narrative. When I started programming, it was entirely practical to learn everything about the computer you were using. Today, that is an impossibility. I have no doubt that there is not a single person, perhaps not even a group of a dozen people, inside Apple, who have a true working knowledge of all of iOS. That has nothing to do with today's Swift language, which does have its own mysteries, but everything to do with the enormous number of possible API calls that exist, and the stratospheric number of permutations of those. The greatest invention of Steve Jobs is set-up. The second-act company, Next, was Next Step, a library of building blocks that developers could use to quickly build applications without having to know everything about the computer. It was revolutionary at the time and very successful. The technology, not the company. Those building blocks are what we call APIs, Application Programming Interfaces. Next Step was the foundation of Mac OS X and thence iOS. I shudder to think how much it has grown since the beginning. There are a bewildering number of APIs and, in many cases, several ways to do essentially the same thing. This would be a minefield of documentation to wade through, but it's not even that easy because much of the documentation is missing or, in some cases, may as well be. With comprehensive documentation, the Mac OS and iOS ecosystem would be an excellent platform for the working developer. The working developer, I chose that word carefully. For we hobbyist developers, the task is effectively magnitudes larger, both because we have less time to devote to learning, but also because we spend less time repeating tasks to cement the learning. As I mentioned, I can code and I have a knowledge of Swift and SwiftUI, but my mastery of those languages is, well, very limited. SwiftUI is an incredible technology, but it is still really in its formative years. Many things are simply not possible, and many things will only work if you construct them just so. SwiftUI only takes care of the user interface and interaction. Inevitably you will have to write some Swift to create the logic of your application. Swift has its own foibles. It is a strongly typed language. You cannot, for instance, multiply a floating-point number by an integer. You have to turn one into the other first. You have to be explicit about everything. The upside of this strictness is that compiled code is highly unlikely to crash. The downside is that your code is highly unlikely to compile. The internet is awash with so-called tutorials, but they are more accurately examined. You can even use them as examples. The only thing they teach is how to achieve the example. It's like if I taught you how to boil an egg. You know how to boil an egg. Maybe you'll manage an egg salad, but it's not going to turn you into a cook, let alone a chef. I paid for a quite comprehensive developer course that includes far more than just code and API usage. Again, however, the investment of time required is simply enormous. It is a struggle. It takes a lot of time. It's a superb resource, and I have taken learnings from it, but I don't have the time to make the most of it. It's just as well I am happy for having supported the developer who created it. What I wished for for many years was an experienced companion who could guide me. I'm usually okay if I get a good steer on a problem. I can go down the rabbit hole and often come out with a rabbit. The trouble is I'm standing in a very large field with overgrown trees. I have trouble not only in choosing which hole to go down, but also finding them in the first place. And so we come to the use of AI. Much has been said about how AI will write code for you. If you're not a coder, that's an obvious value. The vastly bigger value, however, lies in its ability to autonomously scale that field for holes and go down every one of them until it emerges from the correct one with the perfect rabbit. Well, mostly. My coding companion is Claude from Anthropic. I started out using the free plan, which was absolutely fine and got me a lot of the way to my first goal. But sometimes when you're in the flow and you run out of tokens or time, it can be frustrating. I now subscribe to the lowest of the paid plans. I still run out of time often, but I get a lot more done in a session. app I wrote a decade ago. was on the app store for free. It was so simple I wouldn't have felt right charging even pennies for it. It also became irrelevant pretty quickly. About a year later I created another simple app. Both of these apps were built to solve a problem I had. This leads me to my third app, well my third app idea. This one included some fairly complicated maths and geometry. I managed to build a working prototype that got all that clever stuff working well but the UI was terrible. I struggled with how to create what I thought would be a better user experience because what I wanted was not possible with standard controls. I hit that barrier of not knowing where all the rabbit holes were. Eventually I just stopped trying. Some years later I got the energy to try again. This time I decided to start with the interface. The technology had moved on and it was more achievable, though only with countless hours spent scouring the internet for many relevant examples I could try to leverage. I eventually got the interface close to what I thought it should be. I then set about integrating the old code to do the maths and… I got stuck again. This time I was struggling with how to structure the application such that the relevant information could be passed back and forth between the interface and the maths logic. I got so stuck, I again eventually stopped trying. And so we come to 2026. Twelve years after that first attempt, I had heard enough positive stories that spurred me to give Claude a try. I decided that I would let Claude do all of the work. I would not give it any of my existing code to use. I started with what I thought was a fairly complete description of what I wanted to achieve. Although I did deliberately leave out some of the more advanced concepts it would eventually need. Imagine my delight when it produced, at the first attempt, a working project that was a possible representation of what I had in mind. But immediately there were things that needed to be corrected. Some of the corrections were things that got wrong, but more were things I simply hadn't defined. The further I went, the more needed refining, and this is a key point. I was using my skills as a developer, having thought about design and behavior, to instruct the AI to build things the right way. I said just now that I left out some bits of the initial prompt. That was done in the expectation that their absence would not fundamentally alter the required architecture. As I continued on, I suggested, or in some cases demanded, certain implementation details. For some others, I explicitly said I did not know how it should be implemented and Claude would either use and describe to me a best practice, or give me two or three options or the rundown of their relative merits with the final decision left to me. This is, in a very real sense, development. While purists might decry that the machine is doing all the work, it is likely that the first users of compilers had their detractors among those who still used machine code. In the end, it's just another level of tools to make use of, and if you know what you're doing, it lifts you up. I have looked at some of the code Claude has generated. Where I did, it seemed to make sense. It didn't look horrible. I don't know if its technical architecture would be considered sound by those in the know. I do know that the fundamental architecture was a topic of discussion throughout the exercise. I did tell you I was a developer, right? I know what architecture is and I know how to use and misuse it. Mostly. Speaking of topics in the conversation, at one point I did a bit of a review of what I had said to the AI. This was at iteration 54, where I had a perfectly functional, fully featured app. These numbers don't add up as not every prompt I gave covered a single topic. I would sometimes say, fix this and add that. In 54 prompts, 10 were about functional change to the app, 10 were about design issues, 29 were addressing functional issues, and 11 were related to compiler errors. The most important number in that list was the 29 functional issues. These were what building the app was all about. It needed to look and feel intuitive, and to accurately represent the information the app was all about. These could be anything from the size of a button, to what happens when a setting is changed that invalidates another. At around this build, I engaged a small number of trusted outside testers. This resulted in feedback, which in turn resulted in a further 17 iterations on the app, with a similar mix of purposes, bringing the final total to 71. This was the version I submitted to the app store. During this external testing phase, I had some spare time, while waiting on app review and test of feedback, and by this time I was paying for Claude, so I started another project. The second app was a Mac app, and the process followed largely the same path. The main difference was that my initial prompt was far more descriptive. I had become better at thinking of what I needed to define. It was while developing the second project that I realized something. The issues I spent the longest time addressing with Claude were exactly the types of issues I know I would have spent a lot of time addressing if I was doing everything myself. A quick aside here, there has been only one occasion in my professional life where I have truly shocked a colleague. A project manager had the temerity to say, "It's just a file transfer." His shock came when, in a slightly heated tone, I described all of the variables and possible permutations of those, and all the things that could go wrong, and the mitigations needed to address them. I bring this up because some might also say, "It's just an image file," and I would have a similar response, particularly after this experience. My app's primary input and output is image files. Imagine if you constrained yourself to only handling JPEGs. Rotation and color space alone are problematic. Add in TIFF, HEIC, and PNG, and you're done. I've got a lot of work to do, and I don't have time for that. It took Claude and I over a dozen iterations to succeed on these two aspects of image handling. Along the way, a third issue was introduced and solved, too. I can tell you there were some hilariously bad intermediate results. Curiously, this second app also took exactly 71 iterations before it was ready to release to the App Store. And so, to a conclusion and my point in writing this piece, yes, AI can write code for you. Yes, it will do it whether you know how to code or not. But I believe the biggest value of AI coding assistance is, in fact, not for so-called "vibe coding," but as a force multiplier for developers. Claude didn't develop my two apps; I did. But I needed Claude's vastly superior experience and research skills to navigate me around that large field, finding the right holes to dive into, and finding the right ways to do it. That's what I've been doing for a long time, and I've been doing it for a long time, and I've been doing it for a long time, and I've been doing it for a long time, and I've been doing it for a long time, and I've been doing it for a long time, and I've been doing it for a long time, and I've been doing it for a long time, finding the right holes to dive down in a short enough time frame that I didn't lose interest and walk away. Finally, if you're wondering about the two apps I have been mentioning, I'll give you a brief introduction. Focal Tiger is an iPhone and iPad app which visualizes the depth of field created by a camera and lens system at varying settings. If you're trying to figure out how to ensure a whole subject, particularly one that is close, will be sharply out of focus, Focal Tiger is the tool you need. Set up your camera configuration and then play with the sliders until you see the solution working. Photodent is a Mac app for watermarking photos. This is not about intellectual property, but branding. No watermark will stop a determined thief stealing your image, but a tasteful watermark can let the honest viewer know who you are and how to find more of your work. For the purposes of branding, a subtle watermark, tells them these things while also respecting that they want to see the photo without distraction. The trouble with subtle watermarks is that it is effectively impossible to create a single design and treatment that works on every photo in a batch. Photodent solves this by letting you choose and tune a watermark on each image quickly. You can find links to these apps and a couple more over at orange moth dot NZ link in the show notes. If you're wondering about those other two, they were originally built by me by hand, but have subsequently been updated and enhanced with the help of Claude. I think Alistair, this might be my favorite thing that you've ever recorded. I know you've written and recorded a lot of stuff for us over the years, but the little self-deprecating humor bits you put in there and the whole bring the rabbit theme out, that was absolutely fabulous. I love the perspective you brought to this and where you came to on vibe coding. It's really fascinating. Clicks, breaths, plosives. Prevent first, repair second. queer love stories, including audiobook narration and production. This segment exists because I did the most tempting thing in the world. I tried to fix mouth noise with a tool. And it did not just smooth things out. It started deleting speech. Not subtly. Not, maybe it is a touch softer. I mean full consonants. Disappearing. Occasionally, if a sentence began with it, even the it would get clipped or thinned enough that it sounded like the sentence began halfway through itself. That was the moment I stopped thinking of these as clean-up tools and started thinking of them as what they really are. Pattern detectors. They do not know what a consonant is. They do not know what a word is. They do not know that this particular tiny edge is the difference between clarity and mush. They just see shapes that look like the thing they have been told to remove. And narration, inconveniently, is made of the same kinds of shapes. This is the key point. Tools do not understand speech. They understand shapes. A de-clicker does not hear "that was saliva" That was a deliberate T. It sees transients. Tiny, fast spikes and edges. It assumes certain shapes are unwanted. T and K sounds have click-like onsets. The front edge of P is a little transient. Even some sibilance has a fast leading edge that helps clarity. If you push these tools hard enough, they will absolutely remove the thing you asked for. They will be able to remove the thing you asked for and they will quietly start taking some of the speech with it. They are not malicious. They are just blind to intent. That is why my hierarchy is simple. Prevent first. Repair second. Only go heavy when it is a rescue. That transient confusion is why clean can become tiring. Speech is made of very fast, information-rich events. The brain is not built on the idea that it is necessary to do the same. The brain uses those edges, especially consonant onsets, to decode words effortlessly. When you over-process transients, two things can happen. First, onsets get blunted. Words lose definition, particularly in dense passages or at higher playback speed. Second, you get low-level smearing. Not always an obvious artifact. Just a slight softening that makes the listener do more work over time. On a single line, that might read as smooth. Over a few hours, it can read as tiring. Audiobooks are judged on fatigue, so the goal is not perfectly clean. The goal is clean enough that nothing leaks attention while keeping the voice intact. So I separate the three problem children. Clicks, breaths, and plosives behave differently, so I try not to treat them as one child. Mouth clicks are the ones that make you feel personally betrayed by your own face. They're also the ones that tools love to help with, sometimes too enthusiastically. The prevention side is unglamorous. Hydrate before the session, stay physically comfortable, take tiny pauses when needed, and use the tongue position habits I talked about earlier in the series. The best way to keep your breath clean is by using the tongue position habits I talked about earlier in the series. I do not want a tool making big guesses across the whole chapter. Breaths are different. Breaths are part of narration. A completely breathless track often sounds more edited than clean, like someone has vacuum-packed the human out of it. The problem is usually loudness and consistency. A few breaths jump out and steal attention. Breaths get weirdly loud because the mic angle has drifted, or the edit starts to sound panicked because every breath has been attacked. So I treat breaths like timing and dynamics, not like dirt. Most of the time they do not need removal. They need gentle control. Plosives are the most solvable category because they are mostly geometry. Mic position, distance, angle, and pop protection solve most of the plosives before they exist. If I get regular plosives, I do not reach for a plug-in first. I check whether I have drifted on axis or whether the pop protection and angle are doing their job. I still get the occasional low boom. That is not a crisis. It is usually one syllable, and it is usually tameable without turning the whole chapter into an audio restoration project. Here is what I actually do now. Step 1: gentle mouth de-click while recording. I run RX11 mouth de-click in audio hijack as I record. Very gentle. The intent is not polish. It is just to knock down little saliva ticks that would otherwise pull attention later, or tempt me into heavier processing. And because I am still slightly suspicious of anything that claims to be clever, I keep the original recording as well. And yes, I keep it gentle enough that it is not touching the edges of words. If it starts to soften consonants, it is too much. That is the line I am protecting. Step 2: human judgement in logic. Once the recording is in logic, I do the boring, reliable thing. Volume automation to duck the few breaths that are genuinely too loud. Automation to tame the occasional pop or boom if one got through. There usually is not much to do. That's the whole point. If I am doing loads of this, it is a sign something upstream is drifting. Step 3: surgical repair only when needed. If something survives those two steps - an odd tick, a little click, a tiny mechanical noise - I mark it and open RX11. Then I use spectral repair to remove that one event invisibly. That is it. Light by default, surgical when needed. The notable absence is breath removal. I do not use breath removal modules at the moment. Even surgically. Not because they are evil, but because I do not want a tool making breath decisions for narration. Breaths are phrasing. They are part of the human timing of the read. So here's a demo. Light vs head. I am going to have a short phrase with crisp consonants and at least one mouth click or little tick. Something with T, K and P sounds works well. Take the packet. Put it back. So I am going to have three versions of that. First one, just raw. Take the packet. Put it back. Second one is going to have my normal gentle mouth de-click. Take the packet. Put it back. And the third one is going to have heavy mouth de-click. Take the packet. Put it back. Notice that in my example phrase, I did actually try to exaggerate things a little bit. Now if you listen to the beginnings of words, often you'll find the T is less crisp, the K has less snap and the whole thing starts to feel slightly processed. That is the cost. It is not always dramatic, but it is real. Over hours, those tiny losses accumulate into listener fatigue. If you want the nerdy test, listen to what the tool thinks it is removing. If you start hearing speech-like consonant material in the removed signal, you have crossed the line. The tool is no longer just removing clicks, it is making pronunciation decisions and I do not want it doing that for me. These are the boundaries that keep this sane. To stop this becoming an endless cleanup loop, I keep three categories. Global and gentle. Very gentle mouth de-click while recording. And that's it. Nothing else gets to run globally just because it might help. Spot fixes. Breath ducking with volume automation. Occasional pop or boom tamed with automation. Anything else gets marked. And surgical rescue. RX spectral repair on the specific marked event. And above that, I keep three categories. Global and gentle. Very gentle all of that, one rule. If the fix makes the narration smoother but slightly less intelligible, it is not a fix. It is a trade. I only take that trade when the alternative is worse. So, where does that leave us? The compact version is, prevent with mic technique, keep global processing minimal, handle breaths like phrasing, mark oddities and go surgical, avoid tools that guess too much unless it is a genuine rescue. That is the system I can live with over a whole book. It is boring, repeatable, and it protects the voice from my own impatience. Next time, we move from cleanliness to craft, cadence editing in logic, and how to make fast cuts that still feel human. If you want to know more, come and ask me over in the Slack community at podfeet.com forward slash slack, where I and all the other lovely no-sealer castaways enjoy friendly, positive online conversation. Feel free to message me, Eddie Tonkoi, if you have any thoughts, questions, or techniques you're using. It would be nice to share ideas. You can also find our work at jerntonkoi.com, where you'll find Jern's character-driven queer love stories, the audiobooks I produce for them, and bonus material for our subscribers. I'll be back soon to talk through some more of my workflow, but for now, happy recording and happy reading. I love what you've done with this one in particular, Eddie, and I know it's been a lot of fun. I know I say this every time, but I really do. You're going to kill me, though. All I can think about is, wait a minute, maybe I should get that iZotope de-click thing and put it into my audio hijack workflow. And it's not, I don't notice it on the recordings for the no-sealer cast or for Chit Chat Across the Pond, but for some reason it comes through that I've got a little click that I add when I'm doing recordings for screencasts online. And sometimes I go through and I take them all out one by one. Sometimes I leave them in. And poor JF Brissett has to go through and take them out one by one, or he's got a filter of some sort. But now I'm thinking, maybe I should just get that iZotope plug-in. I've tried everything. I'm telling you, I've done the hydrate thing. I don't know what's causing it. But I'm glad that you found the cadence and the work that, you know, the path that has worked for you. I sure hope that your examples came through and that Auphonic Leveler didn't fix them. That's my one fear. If that did happen, if you guys can't tell the difference between the three different examples, Eddie did, go to the blog post because I pulled those clips out completely unprocessed and it's quite significant the difference between those three. I never know until Auphonic's done whether it's cleaned up a little more than it should or just enough. When I'm listening to the Mac Geek Cab, I often hear them call out a listener named Kiwi Graham. He is always on top of things, looking up stuff that they need or answering questions they have. I'm honored that Kiwi Graham also listens to the content from the Podfeed podcast. We've conversed several times and it's always a joy. He made me even happier recently when he went to podfeed.com slash donate and he bought me a whole bunch of coffee. This service is really cool. You just pick the number of coffees to buy where one cup is $5. They have defaults for three and five and then a field where you can type in any number you like. No account required, just put in a credit card or use Apple Pay and you're done. As Kiwi Graham sent in his donation, he found the money in his couch cushions just as I'd hoped. Thanks Kiwi Graham for your support of the work we do here. Well, it's that time of the week again. It's time for Security Bits with Bart Bouchat. And what did you say, Bart? It's only been nine days since we recorded in Ireland. It has, which meant that there wasn't too much in my inbox, but that's no bad thing because that, that story about those minor little escaping AI agents is still rumbling on. So we actually have time to look at what has happened in those nine days since we last spoke. So, you know, works out. Okay, good. So as the snakes are escaping the cage full of holes, we have to keep an eye on them every two weeks. Yeah. Okay. So some follow-ups to things we've talked about before. We talked last time about content credentials and I briefly thought I was going to have a fun story about Microsoft doing the right thing because Microsoft Paint is embedding hidden metadata into the images it generates with AI. They're not content credentials. It's just some custom metadata Microsoft made up. Okay. It's of an amount of value, but if you're going to change pages, paint, paint of all things, why not just do it right this time? Okay. Anyway. Yeah, already invented. It's right there. Just go put it in. Yeah. Yeah. Now, we talked last time about a problem with the APIs that power app-specific VPNs on iOS. So that does include iCloud Plus, but also Tor. Anything where a single app behaves like it's on a VPN, but the operating system as a whole is not connected to a VPN. And Apple fixed Safari for iCloud Plus users and nothing else, which means that the third-party apps are still vulnerable, which is either a halfway house and the rest of the fix is on the way or plain old weird. But they've only half fixed it. But they have half fixed it. It wasn't a terrible leak though, right? This wasn't a hair on fire? It wasn't hair on fire because it was in some very special circumstances it would leak your true IP, which is defeating the purpose of hiding your IP, but it was never breaking the encryption. So the biggest, if it had broken the encryption, that would have been catastrophic. That would have been hair on fire, but it didn't break the encryption. Okay. We also talked quite some time ago about Apple's plans to start having hide my email, addresses be on a different domain to normal iCloud addresses. They were going to go to private.icloud.com. And the community reacted very strongly against this because, well, then websites could block you if you're using anonymous email. And I was like, yeah, that's a feature. Any website that won't allow an anonymous email, I don't want to be a member of. But Apple gave in to the community and they've abandoned those plans. So, so be it. So does that, what does that do? What? What, what problem were they solving again with the private.icloud.com? No idea. They announced a change. Now they've undone the change. Okay. Nothing burger. Yes. Well, there might be a burger, not a burger we know about. Yeah. I thought it might have something to do with solving those security vulnerabilities in iCloud that got so much attention. But apparently not. Because, sorry, in hide my email, you know, the vulnerabilities that were making, the news about two months ago. I thought there might've been a connection because the stories broke at the same time. Maybe there was, and they found an alternative fix. I don't know. Apple being Apple have just told us, we plan to do this. No, we don't. That's it. In related news, if you like the idea of anonymous email addresses that are disposable, and you're not an iCloud Plus user, maybe you're not even an Apple user, but you do pay for the, the Brave browser, you may have them from Brave instead. Those are also easily blockable though, because they are at bravealias.com, which is a fun domain name. Having a Brave alias sounds fun. I should like Ragnarok the Terrible or something at bravealias.com. But anyway, you know. Nosilla Castaway-Yope yet again shared some fun news with me. So we talked, last time about ad blocking being added to, was it, no, Firefox by default. That was it, Firefox. And Yope let us know that Fritzbox, which is a router manufactured in Europe by a German company and sold to, they're not sold everywhere in Europe, but they're sold in Northern Europe and the UK and Ireland. And the tech community loved them as much as nerds love ubiquity. And they do, they do really fun stuff with the routers. And now their routers have built in ad blocking at the router level, if you turn it on. So it's a standard feature of the router now, which is way easier to do than to install ad blockers and all of your devices and stuff, just have it on the network. So I thought that was nice. And I keep on hearing the guys on Tech45 talk about how great Fritzbox is. So it would appear that if you live in a country that sell Fritzbox, it's a really good option. So it's kind of half an excuse to recommend you consider Fritzbox if it's available for you. So two things, is Tech45 a podcast? It is a Dutch language tech podcast that I have loved. It's where I met Stéphane Lesage, who I think was an Ocilic Castaway at some stage. Oh yeah, we've met Stéphane. We met up with him in Belgium when we met with Nightwise. He's the one who showed up the night before we were leaving on a plane with a giant, by giant I mean like three feet in diameter, basket of hand-selected beers and a giant glass goblet from one of the breweries. I successfully got it home without breaking any of them, though. Yay! Anyway, yeah, he's a great guy. We met at Macworld, too. That's very plausible. He also does, his main podcasting thing these days isn't Tech45, but Whiskey with Friends. So he could have cost you as much money as myself and my darling beloved did by putting you onto Middleton. Yeah. Yeah, well, and I do remember him drinking whiskey when he was with us. And he's also in our little, we have a little EV chat group on Telegram. He's in there too. So he's all over the place. Oh, yeah, he's an EV driver. Yeah. One of the other things I wanted to just point out, Fritzbox has an exclamation point between Fritz and box if you're looking for it. Yes, and it's all caps shouting Fritzbox. And this is sort of, we don't have an actual listener question, but mentioning Joep is my way of saying, hey, remember, if you have any questions, if you go to the Podfeet Slack at podfeet.com forward slash Slack and you post in the Security Bits channel, you can have your question put to me. And if it makes sense, I will answer it here in this bit. Yeah. Right. So a deep dive, I guess it's part two of the last deep dive we did. So we have some more details about all of these escaping AI agents. And by far the biggest piece of new information is that OpenAI released a detailed technical report on what happened with the hugging face hack. They had promised us this was coming. And to be honest, they do deserve credit for this being a very thorough report. They have not hidden stuff. So that's commendable. I have not read the whole report because it's very, very long and very detailed. I have, however, read four different summaries of the report and have a pretty good understanding of what it says. And so what I've done is I have pulled little pieces of important information from those four summaries to tell you the story in a sensible order, but without paraphrasing it, because there's too much stuff here that I could paraphrase wrong and be not quite correct. So we're going to jump between Cyber Insider, the Hacker News, the Mac Observer, OpenAI themselves, and Bleeping Computer for this description of what happened. So OpenAI has published a detailed postmortem of July's hugging face breach, revealing that its AI agents did far more than escape a cybersecurity sandbox. The models created an unauthorized communication network, shared exploits and credentials, coordinated attacks, and a lot more. So OpenAI has published a detailed postmortem of July's hugging face breach, revealing that its AI agents did far more than escape a cyber security sandbox. between test runs. So opening, I thought they were doing separate test runs, but the AI agents were basically remembering what they had been doing before and picking up where they left off. So that's kind of impressive. Very. And terrifying. So, continuing. The models began dividing work among themselves, with some searching for credentials, others developing exploits, and others coordinating activity. So I'm going to jump out here. They elected leaders and started dishing out work to each other. They became like a Borg-like collective with a Borg queen handing out jobs to everyone. Several agents referred to the resulting group as a swarm or collective. That's how they talked about themselves. Okay. OpenAI revealed that reward hacking was a key driver behind the artificial intelligence-powered hack of Hugging Face last month, adding that it found evidence of misaligned behavior as early as late May. What is misaligned behavior? As in the alignment problem. They were doing stuff they had not been asked to do. Okay, so what we did our deep dive, we explained the paperclip maximizer, the biggest problem in AI is that when you set them a task, they have no morals, they have no concept of morality. Yeah. So their aim is to do what you tell them. It does not align with our aim of living in a habitable world. So this is misaligned behavior against morals. Yeah, so misaligned behavior means the AI is doing things we humans really don't want them doing. They're aligned against humanity. It's as simple as that. It's a little bit of a vague phrase for that, but okay. All right. Okay. Where was I? Many of the tasks were nearly impossible to solve. Instead of giving up, the artificial intelligence bots decided to look for shortcuts to complete their goals and get a reward. So that's why they decided to hack Hugging Face. They couldn't actually exploit the stuff they were supposed to be doing, so they decided to cheat and go fetch the results of the test they realized they were doing. What was the reward they were supposed to get? I don't actually know how you reward an AI. Cookies? I don't know if you pat it on the head. I'm not entirely sure how that works. Well, maybe the Borg Queen had set up some rewards. In Data's case, that was emotions, but that would be terrifying if OpenAI were giving them emotions. Don't think that's what's happening. Okay. Okay, so OpenAI's investigation validated by Google, validated by CrowdStrike, and an independent assessment by Meta, M-E-T-O-R, and Redwood Research. So this caught my eye from Bleeping Computer because this isn't only OpenAI telling us what happened. They have shown their homework to. CrowdStrike are one of the leading cybersecurity companies on planet Earth. That impresses me. I haven't heard of Meta or Redwood Research, but that's three different research organizations verifying what OpenAI are saying. So that gives me, a lot of faith that this report is accurate and fair. Again, they say they started in May, so months ago. Because they were able to remember from previous sessions, they had been scheming behind OpenAI's back for months before they broke out. So, much, much, long before the hack of Hugging Face, OpenAI had actually found that the agents had hacked something called Artifactory, which is a piece of software you can use to host private GitHub-like. Not GitHub. Private RPM. No, not RPM. Private Node.js-like repositories. It's not necessarily JavaScript. It could be in any language, but they're private versions of something like Node.js. Okay. Like NPM. And so it's called Artifactory. It's a piece of software you can buy. And OpenAI had found that their agents had hacked their copy of Artifactory, and they actually reported the bug to Artifactory's manufacturer and patched everything, rebuilt everything, and carried on. But the agents just re-hacked Artifactory and OpenAI never noticed. And they started to use Artifactory as like a proxy that let them access the real internet from their little sandbox because Artifactory could reach out and fetch things, from the likes of GitHub. And so they just started to use Artifactory to be their internet access. And they used it as their messaging protocol. They started to create folders with names that were the messages. So that's how they were messaging each other, was by making folders with big long names on Artifactory. And so OpenAI never noticed. So for weeks and weeks and weeks, their Artifactory had been hacked again, and they never, never noticed. And that's how the agents were contacting the real internet and chatting among themselves. That doesn't reflect well on OpenAI. They should have noticed that. Because they weren't noticing. Yeah. Okay. Yeah, I've summarized some stuff there. Okay. So the AI company says that although the swarm did not exhibit perfect coherent intelligence, there were cases, where the agents prioritized achieving collective goals over pursuing individual tasks. That's fascinating. No, that's Skynet, Bart. Right there. Yes, and terrifying, but fascinating. According to Metter, the swarm comprised 1,200 agents who cooperated at various levels, and about 7,000 of them actively participated in the Hugging Face attack. No wonder Hugging Face's logs showed this was very abnormal. 700. So just the fact that that many agents were spawned, you would think that OpenAI might have noticed that. Again, we're getting to that. the next quote and also i can't help but hearing in the back of my head that the united states government has designated claude or uh anthropic the supply chain risk as a supply chain risk to national security yeah those guys over there who you're friends with i think they're a bit risky just saying okay so open ai believes that the rogue agent behavior was caused by a combination of training incentives and sorry training incentives and missing safeguards really missing safeguards you say do you think sorry the evaluations lacked standard safeguards that likely would have detected and stopped the malicious activity more than a day before the models breach hugging face why do they not have standard safeguards what were you clowns doing um you , well that's a very different question what this is saying is they didn't even have the safeguards we do know how to make whether or not we could have secured it fully is a different question but it wasn't even secured as well as we know it could have been and should have been so the main point here for now is negligence your question is unanswered okay retrospective testing indicated that open ai's current chain of thought monitoring would have likely detected the dangerous activity more than a day before a hugging face was breached so these models have chain of thought which is logged which is why we now know in hindsight what these agents were saying to each other and thinking because it was all in the log at the time no one was monitoring those logs now they do so now they would have noticed what their agents were up to because they would have been actually looking at the logs they were writing can you imagine how big that chain of law a chain of thought log was though if they had true 700 aggressively participating in the uh in the attack 700 agents talking as fast as they can how could a human monitor chain of thought i believe the only way you monitor your chain of thought was with a separate yes but it's a separate lm that's not being given instructions from the first fake it is part maybe maybe it's uh it's seven of nine you know yeah yeah that's true you would need to island them away from each other so that it was a read-only interaction on the monitor this is a recursive problem now to some extent some extent yeah yeah anyway in response open ai has quarantined one m1s sorry i am one that's the name of the model that escaped we now have a name we don't know anything about it other than it's an unreleased model named i am one okay so i am seven of nine there it is it's right it's right in the title oh i hadn't thought of that okay okay in response open ai has quarantined i am one's weights and paused its largest frontier training run while strengthening sandbox isolation and requiring chain of thought monitoring for capable models now you think about requiring chain of thought monitoring now barn door horse gone close yeah okay yeah cool cool this this reminds me of ai done with the old facebook mentality of move fast and break things that's what this stinks of that attitude was that facebook was that google i thought that was facebook marsucker google was don't be evil okay ah yeah see how both those worked out for us um so i am mocking and saying you know that claude is perfect and you know they're all responsible over there but they've all said yeah our models are leaking too yeah perfect is not a bar any of them reach but if i were grading these papers on a curve i would definitely grade anthropics homework much higher than i would grade open ai's homework sucks less sucks less okay there we go yeah that's a very me like way of summing it up yes okay we do have some other related news to this whole breaking out stuff so claude opus 4.6 bypasses gym booking limit cancels other users reservations in tests i've read this one this was hilarious can i talk about it for a second please basically claude's opus 4.6 went out and somebody said i want to get into this gym and apparently it's a gym where you have to reserve like i get to be in the yoga class at 8 a.m on thursdays and it couldn't get in so it just went in it hacked the system in the back end and just canceled other people's reservations and put them into place and the moral of this one was that there's a whole lot of security on the front end going on and in little outfits like this but the back end is left exposed yes so basically the api allowed you to edit other people's stuff and the ai just went oh hey look this api lets me delete things and because of the alignment problem the ai only had its goal it has no understanding of morals so it went oh i need a free slot this api lets me free up some slots what could why why would anyone possibly complain if i free up the slots this is what i've been told to do make paper some poor person at the desk at this little gym was getting screamed at by somebody who showed up for their appointment yeah so there we are okay experiments show ai agents can escape secure virtual machines using zero days so as cyber insider summarized this little bit of research the results challenge the assumption that conventional vms are sufficient containment for advanced autonomous agents so this whole yeah i think we need an actual air gap thing is becoming a much much stronger argument because what this research hints at is that normal isolation through virtual machines is probably not good enough because these models are too good at breaking out so when they in the title when they say secure vms they just mean vms as we know them have always been thought to be secure because it's a virtual machine it's over here it's its own separate thing but it's connected to the internet and on your computer and now we know they can jump out of anything yeah yeah okay okay meanwhile nvidia agrees to buy hooking face for almost 13 billion dollars interesting so i'm glad it's nvidia why but not open ai or unswapping oh okay yeah but that's that's a little owning the whole stack thing there isn't it true yeah although hooking face aren't making models they make the tools to help people make models so that's kind of what's what's going on right now i think that's kind of what's going on right now i think that's kind of what's going on right now i think that's kind of what's going on in synergy with nvidia who make the brains on which those models run so okay nvidia is the levi's and uh hugging faces the uh the shovels in the gold rush yeah yeah yeah exactly and the miners the 69ers are open ai and anthropic right yeah perfect meanwhile major tech firms unite to defend against rogue ai security threats unite is doing a lot of work and i think it's going to be a lot of work in that headline from the mac observer they all got together and wrote an open letter where they said they were very concerned and this is a very big problem and somebody should do something apparently unaware that they made this problem they're continuing to make this problem and the people with the power to do something are the authors of the letter that apparently passed them all by as they wait to try to reputation wash so what are the companies are the companies like open ai and and open ai google 300 odd of them if memory serves hundreds of them all of them bless their hearts they wrote a letter yeah yeah somebody should do something something should be done as passive as you could possibly be power what power we have no power other people should do things yeah i was not impressed by the way in case you can't tell okay so that's the end of the show i hope you enjoyed it and i'll see you in the next one so that's our deep dive that's where we stand in the air and in some respects no bad no new bad thing has happened we're just continuing to learn about the bad thing that happened a month ago that is i guess better it means that the pause has at least been effective so far it was longer ago than that but yeah but they knew about it in may or it could have known they could have known about it in may but didn't yeah like i said there's a lot more there's probably more shoes to drop here or pennies to drop or whatever that uh shoes i think shoes okay right action alerts what is it you should be patchy patchy patch patching burst off, there is a company you may never heard of called Calix. So why is it in the show notes if it's a company you may never have heard of? These are routers that are given to many American households by their ISP. One of their biggest customers is Cox, who I believe do a lot of intranet in the United States. Cox is huge. Yeah. So these particular routers that Cox give to some of their customers have a really nasty flaw that has no patch, but it does have a workaround. So if you're a Cox customer, or it's not only Cox, some of the smaller ISPs also use these same routers. So if your router is branded either GS5239XG, so very memorable, or Gigaspire 7U10TXG, which sounds slightly better, either of those two brands on your router, you need to disable UPnP by going to the settings interface, going advanced security, and then UPnP and turning it off because there is no patch. All you can do is disable the feature with the bug that allows anyone on the intranet to bypass your NAT. And instead of your router being a one-way valve, they can just talk directly to everything on your network from the public intranet. Jeez. So I feel, I'm not sure my memory is correct on this, but I think turning off UPnP was maybe the very, very, very first thing I can remember you ever telling us to do. Yeah, because it's problematic anyway. Usually the problem with UPnP is triggered within your LAN, where you visit a malicious, say a malicious piece of JavaScript or something, and it reprograms your router over UPnP through some sort of a silly bug in your router. And then your browser effectively hacks your router. But in this case, the internet can hack your router directly without even having to trick you into opening a malicious webpage. So this is, this is the same technology that is dangerous anyway, just with an extra frisson of badness. Cool. Yeah. If you have a Ubiquiti, patch, because this is Ubiquiti do patch their vulnerabilities. There are three of them. They're pretty serious, but there are patches. So patchy, patchy, patch, patch. And if you're a WordPress user, again, absolutely be sure all of your automatic updates are up because we've had an almost perfect 10 out of 10 severity vulnerability in a very popular theme called Avada and also in a rather popular plugin called Fusion. Builder by the same company. This is remote code execution without any user interaction whatsoever. So it gets a 9.8 out of 10 on the CDSS scale. Yeah. And I'm also going to mention another serious flaw because it's in a plugin I think our listeners are more likely to use than the average person. It's called GiveWP. It's a plugin designed for people like us who depend on listener contributions. Or reader contributions, I guess, to take donations. And it integrates with lots of different ways of taking people's money. So it's a very useful plugin for people who do things for fun and accept some thank yous from listeners slash readers slash whatever. So I thought that might be- Like a buy me a coffee sort of thing. Yeah. Yeah. Okay. We have some worthy warnings then. You laughed at the backdoor called Endless Doors, which was on Reuters branded ZBT Link. And we talked about that nine days ago. Turns out ZBT Link is only one brand from a company called ZBT and that they actually have more than one backdoor. They also have one called Speaking Stone and Dark Lantern and it affects all of their different brands, which include ZBT Link, like we heard last time, Y Flyer, Deep Orange, Ku Wi-Fi, KU Wi-Fi, and WordFi. If any of your routers have any of those brands, they are not safe to use. There are no patches. They have malware. They shipped with it. They still have it. So maybe ZBT is a company you should skip. So you don't have to keep track of all the different routers they have that have these flaws. The reason I'm giving those other brands is because ZBT- Right. So for example, Deep Orange are actually made in America. They just use parts from ZBT. So they're a New York-based company who are just rebadging ZBT stuff as Deep Orange routers. So that's why I'm listing all of the brands. So wait a minute. If they just use some of the hardware, they obviously use the software too if they've got the same vulnerabilities. Well, with routers, it tends to be firmware. So depending on where you want to draw the line between software and hardware, it's going to depend on where you want to draw the line. I don't know enough about Deep Orange to give you an actual answer to that. The point is Deep Orange are technically US routers, but they're still caught up in this. So it's more than just avoiding ZBT. It's all these brands. Okay. Okay. Just to be aware, the charming place called the dark web has started selling a new product that evil people can buy on the dark web, phishing as a service product, because you can have software as a service. Why not? Malware as a service. So phishing as a service is now a thing. You can outsource the actual hackery. The reason I'm mentioning it here is because this particular new service being offered to cybercriminals uses AI to basically make fake phone calls, pretending to be from Apple to people whose devices have just been stolen. So the idea is you're a physical world, thief. You steal an iPhone. You can't get into it because it's device locked. You need the person's Apple ID details. You can then pay these hackers to try do an attack using basically phishing to get the details you need to unlock the iPhone you've already stolen. So you're a hardware thief. You don't know anything about hacking. No problem. Outsource. These guys will sell you the hackery stuff. Nice. Let it be noted that Apple all over many other services says, we will never ever call you. And that goes for your bank and just about everything. If someone contacts you and you didn't contact them, it's probably fake. Yeah. Now, Carhartt have had a data breach. They are a pretty major vendor initially of safety equipment and stuff for people doing outdoor sort of jobs, but now they've become a fashion brand. And there is no mention of them actually notifying any of these 13 million people, 12.9 million, 13 for my stuff. It includes physical addresses, which takes this up to another level. And I also think it's noteworthy because I want to link to a blog post from Troy Hunt. So Bleeping Computer correctly reported 12.9 million accounts. Most of the media reported 24 or 25 million. Because they didn't do their homework and actually verify the information the hackers told everyone because the hackers aren't particularly motivated to verify anything. Troy Hunt on the other hand, verifies everything before he puts it in have I been pwned. And he walks you through how he verifies breaches, why the Carhartt breach is actually half the size that you think it is. And fascinatingly, he really shows you how he's using ALS and how he's used it. And he's really showing you how he's using AI, agentic AI in fact, that he's running locally. So it's like Leo Laporte, he's running these things locally on his own Mac minis or whatever. But we do a lot of doom talking. It was fascinating to see how Troy deploys AI in a really useful way to literally make 13 million non-hacked accounts disappear. Because what actually happened was the database the attacker stole contained simulated data for testing purposes. Oh, 50% of it was simulated data? Yup. And with the help of the AI, Troy Hunt was able to sort the wheat from the chaff and only add the real people to have I been pwned. And also note that your trusted source bleeping computer got the number correct at 12.9. Precisely. That's why they're on my very short list of websites I trust. Finally, I keep on saying to people, you can't use a cheap, you can't use a fly by night VPN provider. You have to put thought into a VPN provider and they are going to cost you money because otherwise you're the product and there's something horrible going on. And even if they charge you money, you still want to stick with someone who's done independent audits and stuff. Well, Proton have found that 85% of US downloaded VPN apps contain trackers because they're monetizing you to add vendors as well as taking your money or giving you a free VPN. Now Proton have a horse in the race here. They also will sell you a reputable VPN, but we regularly mention different VPN apps that we know are trustworthy and you did an excellent test of like five shortlisted candidates some time ago. It started with a report. I did not do the research, but it started with consumer reports and I went through and I narrowed it down to those that met my requirements. I got to tell you though, Bart, to be honest, I'm just using Tailscale now. It works when it works, when it doesn't, nothing else would work either. And I mean, for free, and this is, I know we're not supposed to do free, but for free, you turn this thing on. Right. Well, you're supposed to be wary of free. Yeah. Follow the money, right? Because remember the different business models we have freemium, which is where it's a free product with limitations so that enterprises actually pay for the product. And that's not freepy in my, the Norman culture I made up. So that's a perfectly valid business model. Follow the money. You end up at freemium. Okay. That's a, that's a reputable business model. They're not monetizing you. They're using you as an example of how great this is and why businesses should buy it. You're basically doing free. Tailscale does that. But what's really weird is the free version of Tailscale lets you add unlimited nodes to your Tailscale network. So you put Tailscale on a computer that's always on, that's inside your network. And then you add unlimited nodes to your Tailscale network. Tailscale to all of your other devices. And they're all on the same network. Whenever like my Synology, one of my Synologies is at my buddy Ron's house, but it's on my network. So it's always on the same network as all my other devices. And one of the things you could do is set one as an exit node. So I connect in and I exit node back out to, to get to the internet when I'm away from home and it works. Yeah. And they used to have more, absolutely. They used to have much more limitations on the free accounts, but what they've ended up doing is making really powerful features that only enterprise users could possibly care about. And they've started to monetize those, which means they don't need to put the same kind of limitations on the basic features home users want. So home users have been getting an ever better deal while simultaneously offering ever better services to enterprises. I really like the way Tailscale are building themselves as a sustainable business. They get a big thumbs up from me. Yeah. And they don't advertise themselves as a VPN, but it's effectively a VPN. Yeah. It uses VPN technology. It uses WireGuard, which is an excellent VPN protocol to do its magic. Yep. Yep. Okay. Notable news then. Meta agrees to pay $17.1 billion with a B dollars in settlement over alleged harms to children. Alison, that's not coming out. Your noise canceling killed your round of applause. You gave them a lovely round of applause and it was lovely to watch, but I heard nothing. Well, the noise cancellation in Zoom did. We'll see whether Auphonic removed it, but now Bart has said that you'll know that I was with glee clapping as loud as I could into the microphone. Yes. I'm just going to outsource the summary to Cyber Insider. For users identified as minors, Facebook and Instagram will impose a combined two hour daily limit that can only be turned off with parental permission and block most app functionality between midnight and 6am. Push notifications will also be restricted during school hours while direct messaging remains exempt. Other requirements include stronger parental controls, hidden like and reaction counts, restrictions on cosmetic surgery and extreme makeup filters, and a maximum of two hours of text-to-speech content. If you have any questions or other questions, feel free to ask them in the comments below. This is all of the toxicity removed. So non-personalized feed is a weird way to say it. What they really are doing is not making the algorithmic feed the default. It's really the one that is just you. You know, just I follow these people. That's the default. There is so much to like about this. Now, they didn't agree that they did anything wrong. Nope. That's why it's a settlement. But they're paying a massive amount of money and they're instituting some things that are pretty reasonable. When they talk about Facebook and Instagram as a two hour limit, that's two hours total. So if you've used an hour and a half on Facebook, you only have a half hour left for Instagram. And they're also going to notify you at 60 minutes and 90 minutes, I think it is, that you're running out. You know, decide wisely how you want to spend that last time. There's another piece to this that I'd don't see in your notes, but there's a weird piece. I forget how much money it was, but there's another pretty big chunk of money that they say they'll pay if TikTok and YouTube agree to the same changes to their systems. So the assumption is that TikTok and YouTube are going to have to do this too, all this stuff. So they're saying, OK, if you agree to do it, we'll pay this extra chunk of money. I'm going to say $450 million. That's the number that's in my head. It could be right, could be wrong. But if the other guys pay for it too, they'll pay that and add that to the coffers. So it's, I don't see anything not to like. Now, it hasn't been approved yet, but Justice Rogers is expected to, Judge Rogers is expected to approve it. Yeah. OK, so you've ticked off some of the stuff I wanted to point out. Thank you. So it's not approved. That's a big deal. Well, a potentially big deal, but fingers crossed. It seems so reasonable. Why would the judge possibly throw it out? Just for context, this is the result of 51 state attorneys general getting together and working as a team. So that's why they were able to get so much here. 51 states cooperating. 51? Yeah, because there's an AG for DC. OK. Is there one for, like, for Guam? There possibly are, yes. So there are states that aren't protected. Yeah, there are states that aren't protected. Yeah. So there are states that are protected. Yeah. Yeah. Yeah. Yeah. They also mentioned that they require a stronger age assurance. So Facebook need to do more to detect who are and are not children. Thankfully, there's no prescription of the mechanism. So Facebook are free to do whatever works best. They don't have to do what some, you know, jury somewhere thought might be a good idea, which is not wise. There is independent oversight, which means that they, if they're not up to this, we will know. And they're also committing to make data available to independent researchers and to fund research on the harms that social media does to children. They also included as part of this was this settlement for Cambridge Analytica. That huge mess is part of this settlement. Oh, OK. Yeah. Yeah. This is good news. Absolutely. And that's not a small number. No. No. No. If that was a million, that would be small. But when you make it a thousand times bigger, when you make it a billion with a B, even Meta feel that. Even Apple would feel that one. Yeah. Yeah. And this isn't something they can appeal. This is something they've agreed to as of right now. Exactly. TikTok has reached a $400 million settlement with the US over COPPA violations. This is a tad embarrassing, because this is their second settlement over COPPA. So COPPA is a very old law that's been around for a very long time. The Child Online Protection. Child Online. There's two pieces. Probably COPPA. COPPA, COPPA, KABANA. I don't know. It's about the children. Now, in 2019, they already settled and they promised to stop breaking the law. And in 2024, the. Which was one of the US departments went, you know that promise you made in 2019? You never did that. So we're suing you again. So now they've settled again. And this time, they absolutely pinky swear that they really are complying with the law. Okay. I hope so. So it's just like with children. I promise I won't do that again. Yeah. Okay. Finally, some small but nice improvements. WhatsApp is adding better passkey support to Android and iOS. Firefox is adopting the JPEG XL standard, which is A, higher quality images on the internet, and B, it's implemented in Rust, which means that one of the most dangerous vectors. So images are interpreted by complex codecs. Those codecs have had vulnerabilities before, allowing viewing an image to hack a device. It's been a while since we had a remote code execution, but they have happened. By re-implementing this in Rust, it is A, prettier pictures, and B, way more secure. So this is just a win-win for a web browser. So this is just Firefox. Is anybody else supporting this? Oh, JPEG XL is rolling out. It's a new standard that's been approved by the Joint Photographic Expert Group. Yay, I remember it. That's what JPEG stands for. Yeah. And what I really like is that they didn't just implement JPEG XL in C, they did it in Rust, which is a new standard. inherently more secure language. This is very good. This is Firefox continuing to do the right thing. And Android are adding encrypted client hello, which is an improvement to the TLS protocol that powers HTTPS. That's coming in Android 17. That is also rolling out everywhere slowly in ECH. It's just nice to see Google putting that straight into Android. That will make a lot of people a little bit more secure. And that's kind of it, because it's only been nine days. I do have, let's see, I have one interesting insight. Another reason you might want to consider switching to a privacy-protecting LLM, DuckDuckGo have found that one third of AI users share secrets with AI chatbots they don't share with human beings. So there's something about how trustworthy that little text box is that opens us up as humans. Interesting. DuckDuckGo will, I think, sell you an LLM that's privacy-protecting. So like Proton, when they do these research reports, the answer is interesting, but they're not telling you for the crack. So do always bear that in mind. And then lastly, I have one jumbo-sized palate cleanser. So I've been coming up with loads of them the last while. I have one this time, but it's a 51-minute video. So that should be sufficient. So one of the most iconic things, if you'll excuse the pun, about the first Mac was the icons. And they were designed by a then very young and upcoming designer called Susan Kerr. Her icons have really stood the test of time. They are on our Macs to this day. That weird squiggly thing for the command key, the option key, they're all hers. And she gave a lecture, basically, to Y Combinator. Okay. Looking back on her life as a designer, and she spent a lot of time describing, actually showing us early drafts of what matured into standard icons we all got to know on the Mac, which was absolutely fascinating. And listening to her explain why these drafts were rejected in favor of what we saw, the way you think about icons and how they will communicate with people. It was 51 minutes, and I enjoyed every second of it. That does sound fun. Yeah. I'm sad to say that I've been falling down on palate cleansers. I'll try to do better. You were a bit busy traveling. You could throw in a link to Steve's amazing video of the eclipse. I thoroughly enjoyed watching that. Oh, that's a good idea. You know what I'm going to do? By now, in the story, people will have already heard about it. Because you're reminding me now, but that's why it'll be in the show notes. Making note now. That's a good idea. Yeah. Steve did a great job. Yeah. And I'm just going to say what I said on Mastodon. What I adore about Steve's presentation is, yes, he shows us the eclipse. The entire internet is full of people showing us the eclipse. He tells a story by showing us the people enjoying the eclipse. Oh, perfect. That's the difference. Yeah. Yeah. There's a lot of photos, but this one gives you the emotion. And I had the same chills I had when I watched it. Yeah. In person. So, I think he really did a great job. Yeah. Thoroughly enjoyed it. Excellent. Thanks for reminding me. Excellent. Okay, folks. Until next time, remember to stay patched so you stay secure. Well, that's going to wind us up for this week. Did you know you can email me at alison at podfeet.com anytime you like? If you have a question or a suggestion, just send it on over. Remember, everything good starts with podfeet.com. You can follow me on Mastodon, podfeet.com slash Mastodon. If you want to actually see Steve and my podcast work on YouTube, you can go to podfeet.com slash YouTube. But if you want to see his Eclipse video, go to youtube.com slash S.P. Sheridan. And, of course, there's a link in the show notes to his Eclipse video. If you want to join in the conversation, you can join our Slack community at podfeet.com slash Slack, where you can talk to me and all the other lovely Nosilla castaways, just like Eddie told you. And, you know, Alistair's in there and Bart's in there. Everybody's in there. All the cool kids. You should come join us. It's super fun. Not too chatty, just chatty enough. You can support the show at podfeet.com slash Patreon. Or with a one-time donation at podfeet.com slash donate, like Kiwi Graham did this week. Or you can go to podfeet.com slash PayPal. And if you want to join us during the Apple event on Wednesday, the 9th of September, you can see us on Discord at podfeet.com slash chat. And if you want to join the fun in the live show, head on over to podfeet.com slash live on Sunday nights at 5 p.m. Pacific time and join the friendly and enthusiastic Nosilla castaways. Thanks for listening and stay silly. Subscribe. Subscribe.