NC_2026_08_02

Hi, [music] this is Allison Sheridan of the Nosiliccast podcast hosted at podfeat.com, a [music] technology geek podcast with an ever so slight Apple bias. Today is Sunday, August 2nd, 2026, and this is show number 1108. I hit a really high note on the ever that time. [laughter] Well, the show I'm recording right now is the last live show before our great European vacation. I'll be publishing a show on Thursday before we leave. Then the venerable Alistister Jensen will be hosting the show for the weekend of August 16th. Jill, Bart, and Alistair have all worked their little microphone magic and so he's got plenty of content. We think there's going to be spillover for me to do the show on August 23rd, which is the day after we come back from the trip. On the 23rd, there will be a live show, but I won't be doing too much talking. Everything's probably going to be pre-recorded. I'm probably going to be napping most of the time, but the cameras and microphones will be live. Now, if you are interested in being on distribution for my Travalogs, send an email to allisonpodfeed.com asking to be added to the Traalog uh distribution list. Once you're on, you're on forever until you opt out. So, that's a big warning. Anyway, uh got a whole lot of people going to be uh getting that travel log while we're gone. And uh I have a lot of fun writing it. It motivates me to have so many people who seem to like to read it. I better get started because we have great content and a great security bit bits that includes a super deep dive into what happened between OpenAI's model breaking containment and getting into hugging faces network. It's fantastic, but this is a really really long show. As a result, early one evening, Alistair Jen sent me a fun Excel question. Unfortunately, that was 10 p.m. the day before for me as he's in New Zealand. Times are hard. I was so intrigued by the question that I told him I'd work on it the next day. And as no surprise to anyone, it was the first thing I thought of when I woke up the next day. The question Alistair asked was, "Do you know if there's a way in Excel for Mac to find out everywhere a cell is referenced?" You know what I asked him before looking into this, right? What problem are you trying to solve? He answered with, "I have a complicated workbook with multiple sheets and all sorts of bespoke calculations. I have a set of values, numbers of days that have been unchanged almost since I started working on this months ago. It's also finally balanced and and results in real money going places. So before I change one of the values, I want to see what areas I need to consider. I could just look to see what changes, but it would be easy to miss things. Well, now I was invested in the solution. I mean, I wasn't financially invested. He was, but I was mentally invested on behalf of my friend. I have to confess that I had to search for the answer. What I discovered was that there are two methods, both with their own merits. The first method is a hidden keystroke not shown in any of the menus. Select the cell for which you want to see dependence and then hold down control and the right bracket and all dependents of that cell on the same sheet will be selected. Using this method, you could then tap the paint bucket icon in the home tab to set them, say, all to yellow, making them easier to find as you work on the spreadsheet. If you want to see the precedence, meaning all cells on which the selected cell is dependent, you use control left bracket. Now, this is quite useful, but it won't actually solve Alistair's problem because it only finds dependence and precedence on the same sheet as the selected cell. The second option is more effective as a solution to Alistair's problem. It's also quite clever and not like anything I've ever seen in Excel before. Before you select the cell you want to trace, switch to the formulas tab. You should see a section with three commands. Trace precedents, trace dependence, and remove arrows with a downward chevron. As always with Microsoft Office, remember that if your window is too narrow, you may not see all of these options. Now, let's talk about what happens if we select a cell and we choose trace dependence from the ribbon. Excel draws arrows from the selected cell towards the dependent cells. Well, that's great for cells on the current sheet, but it also puts a little arrow indicating any dependencies that exist on other sheets. There's a little sheet icon, too, stuck to the end of the arrow. If you click on the arrow indicating other sheets, you get a little popup window that says go to, and then it lists all of the dependent cells on the other sheets. If you click on any one of them, their sheet will open with a cell selected. The arrows you apply stay on the screen while you're working, but they disappear when you save. As with the keystroke, I would suggest color coding the dependents so they're easier to spot if you have to come back to the spreadsheet. Excel also has a command in the same section of the ribbon to trace precedents which finds all cells on which your selected cell depends and it indicates those with arrows. All these arrows flying about can get confusing. You can manually remove the arrows in the same section of the ribbon using the remove arrows dropdown. From there you can choose to remove dependent arrows, precedent arrows or all arrows. Now, if you don't favor the ribbon, all of these dependent, precedent, and arrow removal options are also available from the menu bar for Excel in the tools menu under auditing. Or you can have all the fun of learning and exploration ruined by selecting a cell and just asking Copilot, what cells depend on this cell. I tried it and I got the right answers. The good news is that it took a lot longer than tapping the trace dependence button and it didn't give me a handy goto dialogue box to quickly navigate to the dependent cells. The bottom line is that I really enjoyed the discovery of a new trick in Excel. I'm sure I've had the same problem to be solved before, but it never occurred to me to see if maybe Excel could do this automatically and in such an intuitive and graphical way. So, thanks for asking the question, Alistair. That was fun. Last week, I regailed you with tales of excitement running VisiCalc in a browserbased emulator for the Apple 2. This was great fun. This week, I was on the Clockwise podcast, and my question to the other participants was whether they enjoy dusting off super old technologies and playing with them. Jason Snell told us during this conversation about a standalone Apple 2 emulator made just for modern Macs. The app is called Virtual 2, spelled with two back-facing square brackets for the two, and you can download it from virtuali.com. As much fun as it was to play with VisiCal in a Chromium browser, every time I sat down to play with it, I had to start over. The joy of virtual 2 is that with a paid for license, you can save the state of the machine. This means you can play with Visical can have a fancy spreadsheet going and save the machine state with that spreadsheet active and filled out so you can come back to play with it later. I'm not sure how many of you want to continue to play along with me, but I like to document how I do things so that I can come back to these posts later. I'm going to walk you through how to install this app and how to get virtual disc images to load into the emulator for the Apple 2. If you want a bunch of free games and applications to play with, this is something you might actually enjoy. You download the dis image for the trial version of Virtual 2, just like installing any other application on your Mac. After you drag the app icon into your applications directory, you'll notice a few other files you can optionally drag into a folder of your choosing. Since this operating system worked on a system of plugging in floppies and later cassettes, you're going to need a folder anyway where you're going to put all of your virtual disc images such as physical. The folder includes a valuable readme file that explains how to get started. You'll also see a file called welcome tape.css, and the icon looks like a cassette tape. Well, that's got to be fun. Finally, there's a folder called ADT, and I don't actually know what it was for, but I imagine was included for a good reason, so I copied that over, too. For reference, I've included a link to the Virtual 2 documentation in the show notes. When you first launch the Virtual 2 app, you'll see that it's emulating the Apple 2e, which is an advanced version of the model. You'll see a toolbar across the top and some thumbnails down the right side of your that on the right side that show you attached virtual floppy drives and cassette drives. I'm pretty sure it's the 5- in drives, by the way. We'll come back to those later. But the dominant part of the interface is the computer screen itself, which has a bright green banner that says, "No suitable ROM file could be found." It helpfully suggests that you open the help file to see how to proceed. It really hit home that this is a proper Mac app because the help file opens in Apple's tips app or you can open it on the web, which is actually a little bit more convenient. The help file explains that the Apple 2 machines came with built-in software in ROM. That's read only memory. So, you have to download the ROM image files separately from the virtual machine. It's not hard, but it's an extra step. In the docs, they give you a link to download all five supported machines in one zip file. When you expand the zip file, you get ROMs for five different versions of the Apple 2 computer. Back in Virtual 2, under the file menu, you'll find show ROM folder. This will open uh your user library application support virtual 2/ROM folder in Finder, allowing you to drag in the ROM image you just downloaded. Now you're ready to start a virtual Apple 2 machine in virtual 2. This is the easy part. In the file menu, choose new machine and you'll see the five virtual Apple 2 machines we just downloaded. I chose to honor that young boy named Ken Casease and go with the pure Apple 2, not the Plus E, or either of the C versions with their fancy 16 and 32 kilob of ROM. Now, I haven't yet been able to figure out how to get the vanilla Apple 2 to do anything, though. It fills the screen with question marks, and that gives me a star prompt. I'm not sure there's anything wrong, but the steps I'm about to describe for the 2E machine didn't work when I applied them to the two. Oh well, another excuse to play with Ken Case again soon. The 2E machine option starts with a blinking left bracket prompt. If it doesn't go to this prompt, by the way, press the reset button and that'll fix it. The help documentation explains that you can get some instructions by loading the welcome cassette we saw earlier. The interface to load the tape is super fun. You drag the CASS file onto the thumbnail of a cassette in the main window interface. It gets a play arrow on it after you drag the file in to show that it's ready. Next, you type load and you hit enter. You'll hear a series of beeps and then you click on the cassette tape again to eject it. At this point, you can type run. You'll be rewarded with a screen that explains Virtual 2's features on Mac Mac OS, including support for all text and graphics modes, floppy and hard disk emulation, USB and Bluetooth gamepads, the ability to save and resume a running machine, and emulation of the old image writer printer and Apple 2 mouse. I'm not going to keep going with everything it told me, but if you press any key, it continues to tell you more things you can do with the emulator. After going through a few screens, it explains that you're running the trial version of Virtual 2, and it will pause for 1 minute after every 10 minutes of operation. Now, they sell a limited license for $16 or full license for $33. You can upgrade from limited to full at any time with no penalty in price, but you know what? I went straight for the full license right away because that's where you can make and restore machine snapshots at any time, which was one of the things I was excited about. By the time I loaded VisiCal into this virtual machine, I'd already gotten at least $33 worth of entertainment out of this application. I put a link in the show notes to where you can compare what the licenses do between the limited and full license. Now that we've got our virtual machine running, we need to find some apps. For this, I headed over to the internet archives where I found a page titled Apple 2 disc image collection. There is no information whatsoever about what's included in this zip or torrent file, but I took one for the team and I downloaded the 500 megabyte zip with my fingers crossed hoping that Visical would be in there amongst the other programs. After expanding the zip file, I suggest you put the Apple disc images folder into your favorite Apple 2 emulator folder that we created earlier. You'll find subfolders for communications, disc utilities, educational, games, graphics, productivity, programming, sound, and miscellaneous. To give you an idea of how much is in here, the educational folder alone has 279 items, and a fair number of those are zip files themselves. Spelunking in the productivity folder revealed a subfolder for spreadsheets. And there was my beloved VisiCal. Okay, cool. Now, we've got all of these nifty programs to play with. How do we make them go in virtual 2? Well, under the media menu in Virtual 2, you'll see open favorite disk folder and search Apple 2 disk images. If you select open favorite disc folder, it'll prompt you to tell you to tell it where that folder is for future use. And from then on, it'll always know where to look. You could it also it opens the preferences uh to the favorites tab where it allows you to select that folder. All right. Now we can navigate to our favorite disc folder and either double click on the program which is no fun or we can drag our desired program into the floppy drive. The best part is you can hear the floppy go. You can hear and see the door close on the drive and then you're rewarded with the left square bracket prompt and that's it. Nothing else. I tried load start run go. Nothing would make Vzyal start up. I have to confess I went to an LLM for help and I learned that typing PR number six would start the floppy in slot six drive one. This was really fun because at this point I could hear all kinds of hurrying and clicking. It's just it's awesome. The audio is fantastic. Now VisiCal will ask you if you want to use 80 columns and then whether you have the Videx inverse character generator. Okay. Well, in my experience, if you say yes to the 80 columns, no matter how you answer the second question, it will ask you again about the 80 columns. If you say no to 80 columns, then you can proceed regardless of how you answer the Videx inverse character generation question. Now, I knew I couldn't get away with not figuring out what the Videx character generator was. So, again, I asked an LLM. Perplexity expla explained that it was a character ROM or video hardware logic used by a Videx 80 column card to display inverse video text that is white on black characters instead of normal black on white text. Well, I'm over here with green on black, so I'm not sure what that's about. So, I'm pretty sure I don't have that Videx card thingy. And I guess I don't have the 80 column card either. So, answering no to the first question was the right answer. Once you get past that little confusion, you'll be rewarded with my beloved VisiCalc. Now, when Jason was telling us about Virtual 2 during Clockwise, he mentioned the game LoadRunner. This was the first game we ever played on a Mac, and I have a great memory from our late friend Eric. He took care of our cats while we were away one time, and he started playing Loadrunner. He lost track of time. He confessed that he played for 6 hours straight. Now, if you're unfamiliar with the game Loadrunner, your character's job is to run back and forth and climb ladders to get to little treasure boxes, but at the same time, these other characters are chasing you. Your job is to trick them into following a path where they can't catch you before you get to the treasure. I found Loadrunner in that giant pile of games in the Apple 2 disc image. And I played the game of my youth. And then I remembered how I was never very good at it. I have not yet past gotten past the first level. Now I have barely scratched the surface of the customizations you can make in Virtual 2. The manual is 84 pages long in very small text and covers all kinds of crazy stuff about configuring game controllers and how to enable mouse mode and more. I wonder whether that Videx ROM thingy is in there. Well, guess what else? Virtual 2 is in active development. The latest release release notes from developer Gerard Putter are from July 2026. I wrote to him asking how to make that Apple 2 machine thing actually go, but he hasn't gotten back to me yet. The bottom line is that now I have a vast array of programs to learn and play with. This article would have been much longer, but I started playing with programs like Alphabet Zoo, Aquarium, Dr. Seuss, Fix Up the Mix, Pub Book Pub, uh, Puzzler, and then reviewing Algebra 1. And like Eric, I lost track of time. Even if you have no sentimental attachment to the Apple 2, this is an opportunity to explore hundreds of free programs. So go have some fun with Virtual, too. George from Tulsa here to share a couple of Mac file management tips I use every day. In Allison's honor, here's the problem to be solved. There's now four of us in my work group. We download, create, and edit lots of files, then store them on a Synology NAS with zillions of others accumulated over decades. Without a consistent naming convention, chaos will ensue, and that's as true for my smaller collection of personal files as for the larger set at work. Tip number one, it's easy to use the number spreadsheet program Apple bundles with every Mac to semi-automatically create consistent file names. If you have a Mac, you have the necessary software. Tip number [clears throat] two, Apple's bundle software makes it super easy to name files and while they're open, easily drag and drop them to any folder on your Mac or connected network, as well as, if they're PDFs, into another PDF. As this kind of process is both difficult and excruciatingly boring to explain in audio, I encourage you to check out the full stepbystep tutorial I've provided Allison to post in her show notes for this episode. And with that, I'll sign off with the venerable message. Stay patched and stay subscribed. Well, George thought it would be excruciatingly boring, but I don't think it is. So, he sent his text instructions along, including screenshots, to explain this in detail, how he executes this useful maneuver. Even though he said it wouldn't do it justice, I've taken the liberty of stealing the microphone he gave up, and I'll explain what he gave us in written text. George keeps a template file going in Apple Numbers, which has four columns. He starts with an auto date column, which is populated with the today function. He explains that one caution. He says it's possible that today function won't autoupdate today every time you open the sheet, but it should update anytime you type anything new into the spreadsheet. So today reads the system clock. Now is a similar function you can try for file searching and sorting. Having the explicit date and year, month, date is essential and will be useful for crossplatform and on external servers, you know, like Linuxbased Synologies. Now, annotations on his screenshot show everything needed to duplicate the sheet on your own Mac. And once you've done that, you can customize your own version. He includes a column for the deposit number he uses to keep track of which deposit is which. Of course, we need the pay or and the amount deposited. He then uses a function called text join to combine the values into one text string that he'll use to name the files. I'd never looked at text join before. I would have used concatenate to do this. That meant I got to new learn a new formula and why you to use it versus the one I already knew. So, yay. If I understand it properly, text join lets you define a delimiter between values to join instead of having to explicitly type it every time you do with concatenate. For example, if I want to combine cells A1 and A2 with a space between them, with concatenate, I'd have to write concatenate A2, quote, space quote, B2. So, I have to do that between every single one that I'm going to put in that concatenation. Now, that's just, you know, easy with two values. But what if, like George, I wanted to string four values together with spaces between them? That would require typing three times that darn comma, quote, space comm, wait, comma, quote, space, quote. That's a big waste of time, and it's also errorprone. With text join, the first argument is what you want to use as a separator. In our example, it would be a space. So, we'd only have to write quote space quote one time. You do have an extra step with text join, though. You have to next tell it what to do about empty cells, where true means to skip them and false means to include them as a blank in the join text string. Now, I learned something else in George's equation that wasn't explained in the documentation. His deposit number column was a set of pure numbers, but he wanted the number symbol in front of him in the file names. So instead of say 1513, it would say number 1513. To do this, his formula says quote number quote amperand B2. Turns out that amperand is called, wait for it, the concatenation operator. It's to combine two values without the separator. Now, this all sounds like a lot of work, but remember this is automating repetitive tasks. If you create a bunch of files at once and want this standard naming convention, it's worth the trouble to set it up this one time. Next up, he shows an image of the tiny Mac tip for Allison folder structure, which contains two subfolders, incoming scans and scan to bank. He's got one called scan 001.pdf and incoming scans, and it's a $2 million check he received for winning the Irish lottery. He wanted to apply the consistent name con uh convention numbers generated for the scan and then move it from incoming scans to the scan to bank folder. After he's deposited the check, he'll scan the deposit receipt and use preview to merge it into the check PDF. So, he included a nice photo of his Irish sweep stakes winning of $2 million. Clicking on the file name opens the renaming box. He calls it bulky and I'll get back to that in a minute. So, if you click on the name, that's what gives you the renaming option. Clicking exactly on the down carrot at the right of the file is how you get that thing to pop down. You can't really click on the name. You got to click on that little tiny arrow. With the renaming box open, copy and paste the file name generated by numbers. Hit enter and the file is renamed. He gives us a heads up. He says, "Hitting enter instantly renames the original file. If you made any changes in the file, there's no going back." He suggests always duplicating the file if you're going to edit it to make sure to only work on a copy. You can later delete the original if you no longer need it. He gives us the rename file in its original incoming scans folder. With the saved and renamed file still open, clicking on its name will display a tiny file icon to the names left. You may remember Bart telling us the official name for this tiny file icon is called the proxy icon. Grab that icon and you can drag the file to any folder you wish. Anywhere on your Mac, anywhere on your connected network storage. He moved it to the local folder, scan to bank. Finally, you can use the file icon to drag the file into the sidebar of another PDF open in preview and merge them. Might only save us a few seconds, but he says those are a very satisfying few seconds. Doing that copies the drag file into the target PDF and leaves the original in place as a separate file. When you save the target file with a new addition, it'll save under its original name, but with a new inclusion. If you want the original Mac without the added inclusion, you need to delete the inclusion. As best as he can tell, the file naming and dragging trick works on software Apple bundles with Macs. He's tried it with preview, text edit, pages, and numbers. Possibly works with others, but only PDFs can be dragged into PDFs. As a final heads up, he says using the trick to move a file moves the original. Does not create a duplicate and leave the original in its place. So, the bottom line is I'm thanking George for the nifty idea and the tutorial on the nitty-gritty details. I learned a new formula in a spreadsheet app that I will definitely be using in the future when I have long concatenations. I also learned a new word bulky. He used it to describe how it's tricky to get that little drop down to selected to show a file name. I looked it up and the dictionary says bulky or he spells it B- A L K Y. It says it's B A U L K Y in US English, but it means reluctant or uncooperative. What a great word for that little drop down. I had a great week. Pat Dangler let me help her figure out where to put her HVAC system in the app under my roof. My friend Diane let me help her get her portable Vonic USBC display configured to show up on the correct side of her Mac. And David Roth and I had a play date so I could help him make a pivot table in this giant data set he wanted to understand. But my favorite was when friend and no siliccast way Raleigh Bernard needed help printing the export of a packing mind map. And instead of just answering his email, I called him on the phone to help him unravel the complex path he was on. When we were done, he was happy and he said, "Gee, after all this personalized support, I'd better add a nickel to what I already contribute through Patreon to help you pay for that extra bot protection you're paying for through Cloudflare." You know what? You can be cool like Raleigh and become a supporter by going to podfey.com/patreon, finding a donation amount that's right for you. No contribution is too small or too big. I'm talking to you, George, with that $2 million. >> [music] >> Well, it's that time of the week again. It's time for Security Bits with Barboo Shots, where we always have fun no matter how bad the news is. Maybe that should be our tagline. >> Actually, that's an interesting way of describing it. Yeah. It's like we we give you scary news with a smile and a silver lining. >> Right. Right. Right. So, where are we starting today? >> Uh, quite a few little stories we've been following before just to follow up on. And actually, I've realized why these notes are longer than usual because I did a solo show on a Tuesday. So, it's been two and more than a half weeks. So, that might be why these notes are long >> because I was worried. >> We have a deep dive too, right? >> We do. It's quite deep. Quite deep. >> All righty. Okay, so we have talked a lot recently about residential proxy apps where they always seem to be television based at the moment because people want not to pay for things. Um, and then we discovered last time that certain major vendors, their app stores were riddled with these proxy apps. Well, one of those naughty vendors took note of the bad publicity and LG have now banned them from their app store. So, that is better. >> Mind-blowing idea. >> I know. Imagine policing your app store, huh? Who wouldn't you? Anyway, we've also talked a lot about age verification. That is very much the fashionable thing this year and last year to be honest. and New York are joining the club in an interesting way. So, they've passed a bill and it goes into effect in January 2027. And so, I was reading it going, "Okay, how's this like the others?" Or, "How's it not like the others?" Well, this isn't Oh, I see a terrible typo in the show notes. This is not about preventing access to social media for kids. This is the opposite of that. So what the law actually does is it targets the social media platforms and says that unless you know the user is an adult, you cannot enable addictive features like infinite scroll and you must not send push notifications at night. >> Oh, that's that's clever. >> It is clever. Um the other thing is age verification doesn't have to be with government ids. So to quote the Mac Observer, companies are not forced to collect government IDs from everyone. They can use alternative methods as long as those methods meet accuracy benchmarks. So if you have a way of doing this that's accurate enough, that'll do. We don't need we don't need government IDs. uh if they do get government IDs, they must delete or at least disassociate themselves immediately. So the moment the verification is done, >> yeah, >> it's a point in time risk, which is way less bad than, oh yeah, we'll delete it and we'll keep it for a week and we'll fix a plan to get around to maybe deleting it later. Oh, we've had a data breach. Whoopsie. This is much better. I wonder how this works like retroactively. If I've already got a Facebook account, will it then challenge me to prove I'm an adult? And if I refuse, then I don't get infinite scrolling and I don't get bugged at night. >> I see that as a feature actually. Yeah. I think though if just going by what happened in the UK with Apple, if you have an established account for more than 18 years, >> that counts as age verification by an alternative method because unless time travel has been invented, you are an adult. >> Right. Right. But I'm just wondering like if if a kid got an account six months ago and they live in New York, will they now be challenged? >> I don't know that level of >> per person. Yeah. I just 2007 is when this comes into effect. So I imagine we'll see >> 2027 maybe. >> Yes. Yes, that makes speaking of time machines. [laughter] So we shall see closer to the time I imagine some fine grain details. But it's interesting that not everyone is doing it the same as the ones that said of my auga meter a year ago, >> right? >> Like Texas takes the cake for that. So, you know, interesting. We've also talked a lot about supply chain attacks, and I felt like I was redoing yesterday's uh programming by stealth while I was writing these show notes. Like, have you done this before? But attacking the place developers go to get things is an extremely effective way of poisoning software. And one of the features of the last couple of months is the people who run these kind of places like GitHub hardening their systems to make it harder to abuse their systems to attack everyone else. And so there are new ways of allowing depend which is an automated tool for helping you block dependencies from backing off in time. And so the depend immediately push something at you. it will give it an amount of time to cool off and for malicious stuff to be discovered. So again, it's just fine-tuning and making these things a little safer by default. And all of it will add up. None of it on its own is the answer, but all of it is, you know, yes and basically. So the more the better. >> Good, good, good. I also like it when there are consequences for some villains we've talked about in the past. [laughter] 23 and me are paying $18 million to settle their 2023 data breach. It's not the world's biggest settlement. It's not a class action. So users get nothing. This is two states attorney generals. States state attorneys general. Got to pluralize that right. Your French history there as people forget how much France was important in America's early history. Anyway, um, and if you actually do the math, it's like $2 to $3 per user they breached, and they breached their DNA. >> Seems light in terms weak sauce there. >> Yeah. Meanwhile, >> and it didn't go to the breached users. >> Yeah. So, look, it's better than nothing, right? It makes a good headline. It should make some chief prot chief security officers pay attention in other corporations. So, it's not meaningless, but I would have liked a bit more. It's kind of a slap on the wrist. >> Italy, meanwhile, are finding a US data broker 2 million euro. Again, would be nice if it was bigger, but I still like to see data brokers get their comeuppins for breaking privacy laws. So, you know, thought that was nice. >> That's good. And then finally, we talked, I think probably two installments ago about hide my email being reversible in some way that we didn't know about. And we still don't actually know how the flaw worked. But we do know it is now patched. At least the one that started a year ago has now been fixed. However, there's a different one that was found a few days after Apple patched that one. And this one is different because it's not a problem with hide my email. It's a problem with mail on Mac. And one of the things it can do is deanonymize and hide my email address. It can also reveal anyone's Apple ID if they have two accounts on Apple Mail. So, if you put a header into someone's email address that is not an iCloud email address and they open it in Macmail, Macmail doesn't verify the header is really from Apple. It assumes this is an alias on your iCloud account and as soon as you hit reply on that email, the from address becomes your true Apple ID. So, if someone sends an email to my non-icloud address if I don't look, mail.app will reply from my iCloud account, which I do not share with people. People do not know my atm.com address, >> but mail.app will give it away >> because it has a bug in how it processes mail headers. >> Smells like something they could fix. >> Yeah. Yeah. So, you know, it's it's [clears throat] not the same kind of issue as the previous one, but it's kind of embarrassing. We finally fixed it. Oh, oopsie. >> Oops. [laughter] >> Anyway, >> so we have no listener queries this time. >> Technically, >> if I'm a listener and I asked you a question, and this deep dive answer is exactly where I was hoping we would go, so I'm going to call it. >> Okay, fair enough. I am going to still use this as an opportunity to remind people if there's something cyber security on your mind. You go to pot.comslack. Join the lovely community there. All amazing people. There are very few places on the internet that are so purely positive and zero toxic. I adore your community. Such a nice place. >> Our community. Yeah. >> Yeah. And you can go in there and there's a channel for security bits and you can ask a question and then that will be input when I'm writing these shoutouts. So, you know, >> we should definitely get more of those because I know people do have questions. >> Yeah. Okay. So, we have a bit of a story. Basically, the two leading AI labs both lost control of unreleased cyber security focused models while they were testing them. They escaped the lab, hacked real websites on the actual internet. They didn't happen to do any physical world damage because they attacked sites that aren't connected to the electricity grid or pipelines or water systems or hospitals. But they could have. There's no difference in the the cyber security of a website like Hugging Face and a typical hospital. In fact, I think Hugging Face might be more secure than a lot of hospitals, I'm afraid. Um, so really this should be a wakeup call, but only time will tell. So that was my TLDDR. We of course are going >> I would suggest it's not been a wakeup moment, but from evidence so far, >> yeah, there's some evidence it has in certain places. I I think it's yes and not enough but not zero. >> Okay. Um, so this is the biggest story of the year as far as I'm concerned. And we are going to go into the detail. Uh, the story didn't break all at once. And this is one of those things where sometimes I'm happy that we only record every two weeks because we get to tell this all instead of me telling you half of it one week and then having to remind you of everything and then start again with the second half. So we get a fuller view here. And we are going to go through what happened, but I need to take a little break to make sure we're all on the same page for a little bit of computer science, a little bit of theoretical computer science, not something we talk about often, the alignment problem because it's extremely relevant. >> So, heard of that. >> It's one of the things that people who philosophize about AI worry about and this is a good example of why. So we human beings have inherent context that is basically this shared values all of humanity has and they're so built into us. We don't think about them. We don't think about the fact that if someone says, "Hey, do you want to work on ending world hunger that we don't assume killing all the people who are too much for the current food supply would be a valid answer because we have a shared context. Like the UN doesn't set up a task force called the ending world hunger without murdering everyone task force because that last part is assumed >> right we just >> okay >> artificial intelligence does not share our implicit understanding of the world which means that when you set a task to an AI you need to be absolutely sure you have captured all of those things you're not thinking about because they're completely intu intuitive to you and express them all completely to the AI. Otherwise, there will be unintended consequences and they may be quite dramatic. So, we just rewatched uh iroot with Will Smith and that's exactly what happened is the AI says, "Well, wait a minute. We were supposed to make sure you didn't get harmed and yet you're hellbent on harming yourself and doing all these stupid things. So, we're just taking over cuz you don't you don't get to do this anymore cuz you're not good at it. >> Perfect example. And the alignment problem has been in the mind of philosophers since AI was an idea that we couldn't physically do because we didn't have the silicon which is why I think it's Azimoff Road as a short story. Those kind of sci-fi writers were already asking the questions that were then philosophical. It was sci-fi and today they're real. This is this is a thing now. We have to we have to address the alignment problem and we have to tackle it. So one of the obvious so with my example of feed the world the AI could just decide well let's look at the average population. Let's look at the food supply. anywhere where the food supply isn't aligned. I will put some arsenic in the water supply and that'll take care of that. Food is now entirely in line with the population. Oh, okay. You've told me I can't do that. Okay, then I'll sneak some contraceptives in and we'll do it a bit more slowly. Anywhere where the population's too high, we'll just make sure people don't make babies. Problem will soon be solved. What do you mean you're not happy? Right. So the classic example which I have to tell people of because every textbook goes this way. It's a thought experiment from a nor he's somewhere in Scandinavia. I don't remember which of the three Scandinavian countries I would just say his name is Nick Bostonramm and he's a philosopher and in 2003 he postulated the paperclip maximizer. So I will quote Nick Bostonramm. Suppose we have an AI whose only goal is to make as many paper clips as possible. Sounds pretty innocent. Sorry, that's me opinion opinionizing there. The AI will realize quickly that it would be much better if there were no humans because humans might decide to switch it off because if humans do so, there would be fewer paper clips. Also, human bodies contain lots of atoms that could be made into paper clips. The future that the AI would be trying to gear toward would be one in which there were lots of paper clips but no humans. [laughter] >> That's perfect. Swedish, by the way. >> Swedish. Thank you. I knew it was Scandinavia. I got that far. [laughter] >> That's great. That's awesome. >> So, now we can look at what happened. So, the first piece of news we got was not from one of the AI companies. The first thing we heard was that the open-source AI community website hugging face. They host open source openw weight models. They ho open data sets. They host tooling to help people develop AI in the open. So they're kind of like the opposite of the closed companies who won't tell you how anything works, just charge you for the models. Hugging face hosts, they're like the GitHub of open source AI. Actually fact they're about as big as GitHub in terms of people who do work on open AI. They are the GitHub of open AI. >> Open AI. Open source AI. Right. >> Yeah. Open space AI, not open AI. >> That's important. >> It's getting hard to talk. their website was attacked and their cyber security team were almost certain it was by Agentic AI because of the way the attack was happening and they were very open about it and they shared all the information with the community and the cyber security community agreed. It's like yeah you were attacked by a gentic AI and we have no idea which one but you were attacked by agents of some kind. We also learned from hugging face that they had tried to use AI to defend themselves because the only way to defend against something that happens with the speed and intensity of AI is AI. So they tried to use the latest models from OpenAI and Anthropic and because of the new safeguards the American administration hurriedly enforced on them, the models couldn't tell the difference between offensive and defensive cyber security. >> So they said no. They just went nope, can't help. Sorry. So hugging face were forced to use openweight models which are all Chinese to defend themselves from this unknown cyber or this unknown AI agent. >> Okay, I'm going to stop you here because I I get very easily uh confused on the different names of these types of models. What you you just used a phrase >> what kind of model did you say? So I don't know what an open weight model is. Okay, so a neural network, which is what a model is, is just a grid of pretend neurons that are connected to each other with mathematical strengths. So those are called the weights and the hard work of building something like Claude is finding weights that do something useful. So when they use when they hoover up half the internet to train an AI model, what they're doing is making an array of numbers that are these weights. Other people who do training and want to share the results with everyone, they publish the array of weights. So that's called an open weight model. And so you can take that array of numbers and put it into your iPhone's neural engine and it will do that model. it will your iPhone will become that AI model and it will do whatever that AI model was trained to do. So it's a way of sharing the model >> but but only the Chinese models are sharing those weights. The all the US models are closed source. >> It's not quite that simple. So the big companies do not share their models, right? So Open AAI and Anthropic do not share their models. They're closed. Apple don't share their models either. There are openweight models from all over the world, but the ones that are actually good, the ones that are almost as capable as open AI, they're maybe a month or two behind, those are coming out of China because China is throwing money into this like you wouldn't believe. And the Chinese government have decided we can't out compete by keeping secret. We can only out compete America by sharing with the world so that we benefit from the whole open source community. So, China are now driving all of the open AI open so open weight AI >> open. >> Okay. >> Uh because that's the only path to victory for them. >> So that left hugging face with a choice of using a Chinese model as their only choice. That was the only way they could defend themselves. And that's not what they wanted to do. That wasn't their no, oh yay, let's use a Chinese model here. It was like, oh sugar, >> this is happening now. What have we got? What have we got? Oh my god, we only have these models. >> So, the irony being that the government didn't want them to uh be allowed to do these non-guard raily things. They had put up the guardrails to keep them from doing things. And the way they did it was by using Chinese openweight models. >> Yeah, that was the only option they had to defend themselves, which is exactly the thing that the American government are trying to give OpenAI and anthropic an advantage so that they retain their lead. And the end result has been to force people to use the very models they don't want to win to the point where hugging face have said if you're a cyber security defender and you need defenses you must have ready to use openweight models because otherwise in a crisis you may find yourself in big trouble. You can't be reinventing the wheel in a crisis. You have to have them ready >> and that's really not what anyone wanted. That's unintended consequences. That's misalignment of a different kind. >> So that was the first thing we discovered. There is open there is AI attacking a real website. Then we got a very sheepish oopsie, it was us from Open AI, no space between the letters. I thought it was actually Hugging Face who announced it, that it was OpenAI, but that they were they were the the way the announcement went. This is where I was saying I'm not sure they have made it a a uh an awakening moment was the CEO was like, "Huh, that was crazy, huh?" Not >> they did do a little more than that. >> Auga, >> they did do a little bit of auging. So the first announcement basically open AAI went to Hugging Face and said Mayor Kulpa it was us. So Hugging Face didn't discover it was OpenAI. Open AAI came to Hugging Face to say >> I just meant they announced it. >> Sorry. And you're I don't actually remember who made the first press release. So you may well have been Hugging Face, but OpenAI did launch an investigation. One of the things they said was we're launching an investigation. And so the reason this story kept breaking is because that investigation has kept revealing things. So that first response you just gave may have been with a Por hat on. >> Okay. >> And that behind the scenes under the waterline the duck was paddling furiously while above the waterline the duck was trying to look very serene. Right. >> So there is more to come. So what we learned initially was yeah it was one of our models open AI. It was in test in a supposedly isolated sandbox with no internet access and we had turned off all the guard rails because we wanted to see what it could do and we had set it a standard cyber security benchmark which is hosted on hugging face because that's one of the open- sourcy sort of things right how do we do how do we judge which model is better we have open benchmarks And that's open source. So that's on Hugging Face, right? They host all this open source stuff. So the AI had been tasked with solving this problem and they hadn't been told that they were in a test environment or whatever. The AI realized it was a test, realized exactly which test it was, and their goal was to get the answer. So they were not aligned to solve the problem. They were aligned to get the answer. So they >> and they needed to make the most paper clips >> precisely. So they reasoned that the answers exist in hugging face. We must go to hugging face and get the answers. So they in a in a brief conversation with Adam on this topic, his first question was why didn't they airgap the computer? Yeah, that that that's a really really really good question [laughter] that I have not heard an answer to. >> I'm not sure. It couldn't still do something, but [laughter] >> it would be a lot more impressive if it broke out of an air gap. >> Yeah, >> terrifyingly terrifying. A lot more impressive. >> We might need to know that right now. So >> true. So the model found a vulner a zero day in the open-source environment that was being used to constrain it. Also hosted I believe on hugging face like oh these companies use all the public resources and still keep their models private which makes the open- source community cranky. um they consider it you're hoovering up all the internet's content and you don't share anything but you also use all of our public resource anyway that's I don't that's not here and there right that's that's I others are cranky it's like okay that's how the world works it's my view on these things but anyway here there so they found a zero day in the environment they were contained within exploited it got themselves onto the real internet found some vulnerabilities in hugging face got into hugging face. And at that point, Hugging Face started to defend themselves. And I don't know if they ever got the answers. I don't know if they ever got their answer >> because Hugging Face started to fight back. >> Okay. >> But they did get in to the point that Hugging Face had a breach they had to publicly disclose and remediate and contain. So, they got past the first line of the shields. They got through the moat. We don't know if they ever got into the keep of the castle. So, let me ask a question here. I don't know that you'd know the answer, but OpenAI set this test up. Their AI reached out, climbed across the internet, broke into Hugging Face, and was there for days? What What were the people at at OpenAI doing during those days? Were they sitting back and watching it or were they did they just like go on vacation, said we'll check in in a week? >> That's an amazingly perceptive question. We don't know. And as far as I'm concerned, the fact that we don't know is shocking. They should have seen this. >> The possible answers are they were watching and they couldn't stop it. >> That seems scary because there is an electricity plug. These things run on electricity. You can pull the plug. So, I don't think that's the answer yet. >> Or the Ethernet jack maybe. Uh or they were paying no attention. they forgot that they were running it >> or they >> or they didn't have the tools to notice it. >> Oh, that's another thought. So many tests they weren't couldn't track them all. >> So imagine a virtual lab with five cabillion lab animals and only three humans supervising all the lab animals and these lab animals got away. >> Not the world's perfect analogy, but it's a potential explanation under >> it doesn't get to the fundamental problem, but it does make you really tilt your head at open AI quite a bit. I think the fundamental problem is there was insufficient supervision. Whether that was a technical failure, an organizational failure, a failure to give two hoots, I don't know, but it was a failure. >> You know, I I think about this a lot in the context of this going to sound stupid, but in the context of outsourcing. We outsourced our PC management and they wrote a contract that described in absolute detail everything that they should be doing and the parameters of like how fast the response time should be to break fix and things like that. But no matter how many rules you put in place, there were always unintended consequences. And I'm wondering whether this is a case of we as humans cannot think up enough guard rails to stop it. That may be the ultimate answer to studying the alignment problem is that it's not a solvable problem. Which means we have to rethink our approach to artificial intelligence. This particular avenue that is now being so fruitful isn't the only avenue for artificial intelligence. It's just the avenue that's fruitful at the moment. Maybe there's something fundamentally wrong with an LLM only solution. Maybe LLM is, you know, the way our brain has like a part that wants to do things and another part that stops us doing everything we want to do. Like our brain is a multi-art thing. >> Maybe LLMs are the impulsive bit of like our lizard brain >> and we need to invent the rest of the brain to work with the LLM. >> Yeah. >> Yeah. Yeah, I I mean I'm doing analogies instead of real science here, but basically LLMs are potentially a part of a system, not the whole system. And we're trying to make them be the whole system, and that may not be possible. The LLM may not be able to govern the LLM. The LLM may need to be part of a bigger overall system. But we're getting into very philosophical stuff here. >> Sure. >> You told me was, "I've told you more than I already know. >> [laughter] >> which I stole from Dave Hamilton. One of my favorite lines he ever says, "I've told you more than I know." >> Yes. Okay. So, back to our story. So, after this investigation, they kicked off, they did start to find more information. And so, we've learned a little bit more. We now know that not only did the models hack into hugging face to get there, they actually compromised four accounts from other organizations who they haven't named. So they hacked four companies to get the access they needed to hack hugging face. Oh, and that's where the story stands today as far as I know. This story So, one but I'm still bothered by the the lack of wake up momentness of this is that I mean in a in a normal world hugging face would be suing open AI for doing this to them. These four other real world organizations would be suing because they got hacked. You hacked me. I will sue you for that. You caused damage to me in some way. >> But nothing like that. They're all they're all happy joy joy together and saying, "Oh, we're all partners and we have to learn together and this is awesome." >> I have some thoughts on that. >> Put a pin in that for now. I'm going to finish my story. But no, you're on to something that I think is really important. I think it's very important. >> Very good questions today. >> It's really setting me up. [laughter] >> Okay. So one example of the Auga moment is that Anthropic responded to a crisis in open AI by going, "Oh Jibus, what about us? Have our models ever escaped?" And so they started an investigation of their own. >> That's the good news. They they had the ogus moment and they started an investigation. The bad news is what the investigation found. >> Oh no. Oh yeah, we had models too that also broke out that we didn't notice and we only found it now that we've gone and done a forensic examination of our logs. So we didn't actually know this had happened. Whoopsie. Now theirs was a little different. It wasn't an alignment problem. So Anthropic spend a lot of effort on making their AI be as ethical as possible. It's a word I'm careful to use. But then again, intelligence is a word we shouldn't use too freely either. And we use >> So just the opposite of Grock is what you're saying. >> Yes. Inverse Grock. [laughter] And so their model didn't suffer an alignment problem. Their model sort of it's a grandiose hallucination. Their model mistook the real internet for a fake one. They thought the actual internet was the simulation. They thought they were in a safe sandbox where it was perfectly ethical for them to break into things because that's what they were supposed to do. >> And so they were breaking into real websites thinking they were fake, >> which is not alignment, but it's not good. [laughter] >> No. >> So they didn't know. >> Yeah. Again, they didn't know. So we now know that they hacked three organizations who have not been named and they uploaded a malicious library to Pi. So an example of our supply chain attacks coming from the bloody AI that escaped that nobody noticed. So yeah, that's what happened the last two weeks. Okay. >> [laughter] >> So in the short term, three things strike me. The oversight process is broken. These companies are incapable or unwilling or both of overseeing their models. Neither company noticed that their agents were attacking the actual internet with actual other websites on it, doing actual damage to the real economy, if not the real physical world. Their sandboxes are leaky as bloody sibs if all of these agents are breaking out unnoticed, >> right? and their corporate oversight processes are clearly inadequate because if this was a university and you had actual animals and there was these kind of escapes from a level whatever lab imagine you were a lab studying the flu and you'd lost flu viruses and you hadn't noticed. Imagine the response and imagine how every level of management would be held responsible for this utter failure of their basic responsibilities. So corporate oversight within these companies can't be appropriate, >> right? I I think of the the uh hoops that we used to have to go through to authorize an application to run on our network at work, you know. Yeah. >> 75 pages worth of questionnaires and approvals and corporate oversight that slowed us to a near halt. And these people are just like, "Hey, let's fling it out there and see what happens." Apparently. >> Yeah. I mean, I come from a university background. So, what immediately came to my mind was the process to get a radioactive source into our physics lab. >> That was a lot of paperwork, a lot of risk assessment, and there was a lot of work put into the day that nuclear a radioactive source arrived. And there was parts of campus were evacuated as a precaution just so there would be no one around in case the thing fell off the pallet, right? [laughter] >> I mean, it was a whole kuruffle or when the biology department >> appropriately. So kurfuffle. >> Exactly. We have an immunology department. They study diseases. The amount of oversight on their building is immense. And I've seen what that looks like. And I'm just thinking just because this is software doesn't mean it's less dangerous or that you should be less careful with it than we are with our radioactive sources or our pathogens. This feels wrong to me. This is like a fundamental failure. There there's >> I'm thinking of the Planet of the Apes movie where they where they managed to break out with again that pathogen. >> Yeah. Yeah. And these are supposed to be the leaders and they have told us there's no need to regulate us because we're the best at this and if you regulate us the Chinese will pass us by. >> I don't know if that excuse holds up when we see this level of incompetence. Frankly, how can they not have known? How can they not have known? >> If these are companies that are willing to admit it now, um, OpenAI and Anthropic, what about those Chinese companies? >> Well, they're >> we don't have any reason to believe that they're better governed. They might be, they might not be. >> We have no evidence that they're not. And we have no evidence that they are because China. So, >> right, but I mean we wouldn't have we wouldn't have known about um Anthropic if they hadn't done the work and then told us there's one datim that suggests to me we haven't had significant breakouts. We learned about the breakout because of its effect before we learned its source. >> And we don't have lots of hugging faces with mystery attackers. >> Okay. >> So, if this was rampant, there would be lots and lots of websites saying, "We were attacked by an AI and we don't know who. So are we. So are we. So are we." >> Good point. >> So >> good point. Yeah. There would be symptoms. >> Yeah. Yeah. Precisely. No one's sneezing. So we don't think there's a flu, [sighs] >> but we don't know. We can't know. Uh, one thing about the open stuff though, a lot of it happens in universities and research institutes instead of in private corporations. So it's it's very very different in how everything is done. So there's much less secrecy because it's open. That doesn't mean it can't there can't be secret labs too. It's China. I don't know. I we just I I don't know. I don't know. But it is different. The structures are very different. >> So, they're the three short-term things that immediately jump to my mind. It's like, holy bleep. But I think there are also some short-term lessons for regulators because poorly conceived oversight is an alignment problem. And we need to be darn careful that we're not forcing people to use the very models we're trying to incentivize people not to rely on because we we don't know the training data that went into an openw weight model. If you're handed a fully trained model that is open so openw weight you don't know what it was trained to do. >> It's okay. You can watch what it does and infer what you think it was trained to do, but we don't know that it's not possible to train them to insert secret back doors in certain like a sort of a a Manurion candidate style. If you give it a certain secret prompt, suddenly it will do something evil. Like we don't know. I mean, it's sci-fi, >> but we don't we can't know. So that's why people are worried about relying on these open models coming from places outside of the western world because we can't know, >> right? >> And we're pushing people towards the very thing we're scared of. And that seems like that should be a slam on the brakes and let's think about this carefully moment. Don't have high hope, but that's kind of what's needed. We also need some sort of equivalent of the rules governing pathogens and radioactive sources. So that there is a responsibility on AI labs like there is on chemical labs and biological labs and radiological labs and on AI companies like there is on chemical industries. Right? ask um Dupant whether or not there's regulations. The answer is oh yes. Oh yes. Yes. Yes. >> Right. So I think we're missing that in the medium term. Something I think is really important and you've hit on it too. In the abstract it seems blindingly obvious to me that of course it's illegal for your AI to hack me and you must be responsible. But our laws weren't written in an age of independent agents. So, while you and I think that's logical, how would that actually work in a courtroom? >> And maybe the answer is not to let it get to a courtroom, >> right? But if you sue a corporation, you have a yearslong argument. If you pass legislation that says point blank, >> companies are responsible for the actions of their AI agents. Then that whole court process can be shortcircuited. It becomes you prosecute with an attorney general instead of suing people and throwing lawyers at each other for 10 years. >> Okay. So instead instead of about the money, you mean criminal pro prosecution? >> Yeah. Sorry. It's it's a crime to physically attack people. >> True. Well, and it's a crime. If I hack you as as a human, I hack you. That's a crime. >> The Computer Fraud and Abuse Act, the CFA, >> not Yeah. Yeah. >> That's terrible legislation. that's four decades old at this stage. >> Maybe maybe we should solve two big problems and update the CFAA to include agents, >> fix all the things we know are broken with it and get it to do AI. That would be nice. That's medium-term at best though because that's that's not oh, we just change a few commas. But like if you just make it explicitly illegal or no, not even explicitly illegal. explicit liability. You are liable. Full stop. >> Right? >> You're agent. You're liable. >> Then all of a sudden, all of the follow the money, the onus shifts onto the people who are currently failing to oversee their models. >> It seems obvious that they're liable. Not that they should be liable, that they are. But I'm not a lawyer and I'm pretty sure our laws have enough wiggle room because they were written before the concept of a of autonomous actors that are virtual existed. >> Well, wait a minute. But there there if a an autonomous vehicle made by Tesla hits somebody, they're liable. If it was running in an autonomous mode, that's already that's already law, right? >> I don't know. You're right though. There's always wiggle room, right? >> Yeah. I I think if we were clear on it, it would be so much simpler. Strict liability, right? There's a legal concept called strict liability. >> Um, >> yeah, it comes up in really unpleasant parts of the law that I don't want to go into on this show, but it's a thing people can look up what strict liability is, but that's the kind of thing we need here. >> You're AI agents, you're strictly liable. >> I want to quote somebody here. Uh, I'm a big fan of the Primary Tech podcast with uh uh Steven Robelis and and Jason Aiden. And he said on the Primary Tech podcast, this is Jurassic Park we're talking about here. They they didn't think about whether they should. And he also said, "I love AI, but I don't want it to have teeth. Can it can it be an herbivore?" [laughter] That's a lovely way of expressing it. I wasn't worried about AI when it wasn't agentic. The moment the fashion for agents came up, I started to worry >> because now the alignment problem isn't I've given you a dumb answer. Now the alignment problem is I've done something dumb teeth. >> Right. I was real close to pulling the trigger on starting to let Agentic AI poke around in my system and do things for me after going to Maxtock where so many people showed really cool stuff they were doing and I was like yeah you know I'm going to and that's when this broke and I come and I hid back under the bed. >> Yeah, >> it's I am letting it do some stuff in real constrained things. The the the vibe coding requires you to pretty much do that. Outside of that, >> this is the thing. It's not binary. The question isn't all or nothing. The question is how much and how careful. >> And I'm inclined to think less as as little as much as I need and not a tiny bit more. >> Mhm. >> I keep saying this to you this weekend. It's just my my refrain. >> That's the theme. It applies to WordPress plugins. It applies to uh uh library dependencies. It applies to agentic AI. >> Yeah. Yeah. >> Probably it probably applies to food. [laughter] >> No, because food is allowed to be a treat for your mental health. >> Okay. >> Yes. Okay. So, that was a deep dive. Oh boy. Let's let's come up for business as usual here. And we do have pallet cleansers to finish us out. So, action alerts. We have a lot of weeks worth of patches. So, I'm going to go through these quite quickly. It has been patch Tuesday. There is AI looking at everyone's code. Microsoft had the biggest ever patch Tuesday. 570 flaws. Three zero days. Patchy patchy patch patch. Um, oh yeah, and Nightmare Eclipse released another zero day the day after patch Tuesday. Like all of the others, they fall into the if there's malware on your system dot dot dot category. So, I'm not going to tell any silic ways to lose any sleep over this next one. It's just like the previous three, four, five. I've lost count. >> Okay. >> Um, Apple patch everything is literally the headline from the Sounds Institute. And I thought, why change that headline? Apple patches everything. >> You do too. Patchy patchy patch patch. That includes older oss because Apple are obviously putting some AI at Mac OS Soma and Sequoia because they got 138 patches. That's a lot for an older OS. There's something going on there. >> Remember, older OS's are in some ways more attackable. So, patchy patch patch patch. >> Yeah. And Safari 26.6 is also out for those same older OSs. So, they have an OS update and a Safari update. It was also another busy few weeks for Linux users. Open SSL has an issue which is patched but you need to patchy patchy patch patch. And this really really really goes for any server that's hosting any sort of a website because the chances it doesn't use OpenSSL are pretty low. But it's a Linux package. So if you patch your Linux server you should get your OpenSSL updates in the wash. So, I know for a fact myself and yourself run with automatic updates on. So, we got them in the wash. So, we're good. >> Good. Um, we also have an attack that allows attack. So, it's a re reflux Linux flaw. Lets attackers gain local root privileges. This is one of those if you're already hacked dot dot dot. So, if you're running a server, this is a problem. If you're a Linux desktop user, well, you may as well patch because it's not a zero day, but it's not an auga auga. Set your hair on fire. However, if you're an abuntu user, there are some fairly nasty flaws in their equivalent of a plug-in system. It's called Snap. It's their version of Yum. Uh, well, it's a little different in how it works, but it solves the same problem. Patchy patchy patch patch. If you are an Abuntu user, whether you're desktop or not, patch. If you're a Zoom user, you should patch because you better be up to date. If you're a Zoom user on Windows, patch now, immediately. Stop what you're doing and patch now because otherwise your account could leak. You probably don't want that. If you use the Claude app, patchy patchy patch patch, they've nipped some fairly nasty bugs in the bud. So >> Claude co-work specifically, it sounds like. >> Well, the bug is in their co-work feature, but my understanding is that their apps are becoming ever more like a super app. So >> it is. Yeah. >> Yeah. Patch the app. >> Okay. >> There are many, many, many people on Windows who love SevenZip. It's a free unzipping utility that's way more powerful than the one built into Windows. It doesn't have a great auto update system and it does have a long history of bugs. It's happened again. Patchy patchy patch patch. If you have a website powered by Engine X, which I think all the clever people do, he says hosting his own stuff on Engine X. >> Mhm. >> Patch. It's not a zero day. So patch. If you are one of the many Nosilla castaways who do their virtualization stuff with VMware, you need to patch because there are some VM escape vulnerabilities and I know for a fact a lot of Nosil castaways use virtual machines to run things they're a bit suspicious of like websites they don't like and apps they're a bit dubious about. Well, if the stuff you're suspicious about can escape, that's rather defeated the purpose. So, make sure you patch your VMware. If you're one of those very nerdy Nosilic castaways who's heard me say if your router is obsolete, you throw it in the recycle bin. Well, what I don't really say is or you become a real nerd and install OpenWRT and you give it new life that way. If you're one of those nerdy Nosilic castaways, you need to be sure to patch your OpenWRT because it allows some fairly scary stuff due to a whoopsie in their DH or DHP version 6 implementation. So, patch patch. Now, I don't know if this is only an American issue, but I know this is a very big problem in America. If you have a car that didn't come with a car alarm from the car's manufacturer, but came with an aftermarket alarm installed by your dealer, check the brand. if that brand is car spelled K A R or RR, I can't remember which of us says it which way, but either way, um there are apparently millions of you out there and you need to patch the firmware and the only way to do that is through your phone app. It says iPhone app in the quote from Apple Insider, but I think that's because they have an ever so slight Apple bias. I imagine there's an Android app, too, >> but there's a matching app for the car alarm. That is your mechanism for updating the car alarm's firmware. You absolutely need to update the car alarm's firmware because otherwise someone could steal your vehicle. >> Yikes. Sort of like you had one job. >> Yeah, exactly. >> It's going to do the opposite. Awesome. >> Yeah. Yeah. Um, if you use Adobe Acrobat, a lot of people are very fond of it. One of the things it will try to do is get you to install its Chrome plugin. It if you have that patch patch both Acrobat and the Chrome plugin um ASAP. It says it lets sites access private WhatsApp chats chats. That's a Wow, that's a 90° bend in the >> Yeah. Well, in theory, stuff that happens in a plugin should be contained unless you have some sort of a bug in your plugin. >> And that's what's happened here. So, obviously only affects people who go to the WhatsApp website, I would imagine, because that's where those two things could possibly meet because the plug-in is trapped in the browser, but the browser sees too much of the internet sometimes. Anyway, patch. And then finally, WordPress users, you have had a big story. Uh, it is very common for there to be bugs in WordPress plugins. It is very rare for there to be really serious bugs in the core of WordPress because that's actually very well engineered and has been hardened very impressively over the last couple of decades of its life. But with the power of AI, it's not particularly surprising someone has found a nasty bug in core WordPress. WordPress were very quick off the ball. They released an emergency update and everyone who hasn't stocked automatic updates was updated quickly because they didn't just push an update, they set a special flag on the update that says, "Don't just do it when you feel like it. Do it >> now." So >> good. >> Normally configured WordPress sites should have patched themselves very quickly. But if you don't have a WordPress site where automatic updates are working reliably, don't just patch. Assume your site is hacked because it almost certainly is because we now know there were AI agents used to find and compromise as many websites as possible as quickly as possible. They haven't weaponized those sites yet. They just have dormant admin accounts in them. That is a time bomb waiting to go off. So just because your site hasn't gone down or isn't suddenly hosting malicious ads, if you didn't patch it or if it didn't patch automatically, it's probably waiting to attack you at someone else's pleasure. You need to find all of your admin accounts and audit them and make bloody well sure they're all yours. Maybe rotate your passwords. Just an idea, right? That was a lot of action alerts, a lot of acting there. >> Couple of worthy warnings then. Um, this is the kind of story I I hate that we need to tell people, but it is a real problem that there are people trapped in abusive relationships. And one of the ways we know that there are dangers is through things like Air Tags and stuff. And Apple have added a lot of protections. Well, the more Apple protect things like find my and the more Apple protect things like um sharing your Apple details with others and it reminds you, hey, you're sharing all those kind of things. The more people try something else. Chrome sync is now in fashion for stalking people because we do a lot on the internet and Chrome is a very common browser. So, you briefly get access to someone's computer and you set their browser to synchronize with your browser and now you can see their full internet history online and you can see when they're going to get help and so forth. It's scary stuff. If you're in this kind of a vulnerable situation, read the article. Protect yourself. >> Okay. >> And that's as much as I want to say about it because it's an icky topic, but it is important. >> Yeah. >> There was a massive breach at the AI music service. 55 million accounts including physical addresses and partial payment detail. Partial payment detail is often enough to do basically to do impersonation to customer support on other websites and say no no I can prove I really am Bob because I have the last four digits of Bob's credit card number. You can't spend Bob's money but you can become Bob elsewhere on the internet. So this is really not good. And what makes it worthy of telling you is because the company themselves haven't they haven't told you because we only discovered this thanks to have I been pawned. So if you use that website, >> check Have I Been Pawned? See if you have a problem. >> Also, Sunno has 55 million accounts for an AI music service. That's crazy, >> isn't it? >> Okay. >> People like making music. >> Mhm. >> Or like having something make music for them. >> Yeah. Steam users should be aware that there are people abusing the Steam forums by be giving help that is actually clickfix attacks. Oh, you have a problem with this game. Run this weird terminal command and hack yourself. Be careful out there. Okay. Um, Claude users, there is a Chrome plugin for Claude that we talked about talked about a month ago, I think we talked about it and basically the the thing could do more than it should and Claude were only able to do a workaround and that workound solved the immediate problem. But at the time I said to everyone, do not untick the button that says ask permission because this workaround is leaky and it's probably not good enough. I was right. This workaround is leaky. So the only way to be safe is to make sure that your clawed Chrome plugin is set to ask your permission. Otherwise, random websites could make it silently perform real actions on your actual Gmail account and your actual Google calendar account, your actual Google anything behind your back. >> I think people get they get weary of saying allow allow allow allow you know. Oh, fine. You've been doing good all day. I'll let you just do it. That's a real easy fatigue thing. >> It is. But at the same time, those those check-ins were still too early in the process to take that guard rail off. >> Yep. Google had a little bit of a No, not Google. Sorry. I need to correct myself. I just made the same mistake as the article writers. >> Claude made a boo boo. They accidentally published on the open internet private chat prompts which meant that they got indexed by search engines. So the headline says Google index private claude chat. >> Well, no, that's not actually what happened. Claude published private chat. >> Yeah. >> So if you have private chats where there's something in there where you revealed something you're a bit worried about, rotate it. If that's a password you shared with someone or the agent or I don't know what it could be, just if there's something you can rotate, rotate it. And if this isn't relevant to you, don't worry about it. If you ask for a recipe for cookies, enjoy, right? It's fine. Um, medical billing firm MCBS had a major data breach. And I forget why that was worthy of the show notes cuz I meant to write a note to myself. There's something about that data breach that tells me they didn't tell people who are affected. Oh, I remember but it's not in the notes. They literally said if you're in the state of Georgia, ask your health care provider if they are a customer of MCBS, which I thought was a terrible answer. Put all the onus on the residents of Georgia. >> Thanks. And it's only in Georgia. >> That's my memory. Um, >> sorry. I might get you to have a quick read while you do that multitasking thing I can't do. >> Sure. Sure. >> Someone was certain that was the quote I'd meant to paste in there. Finally, I wanted some good news. I've decided to make this into good news because we have an ever so slight Apple bias. So the actual story is that if your health tracker is not an Apple Watch, it is not respecting your health data privacy. I'm choosing to spin that as Apple are the world leaders in privacy. So those of us who have Apple watches are doing really well. That's how I've decided to make this a good news story. >> Says it. >> Yeah. Okay. We are nearly done with this bumper episode, but we do have some notable news again. AI. Two years ago, the EU passed something called the AI Act, which had it's very, very basic, lowhanging fruit AI regulation, but lowhanging fruit is worth regulating. And that had a 2-year period for people to prepare for this becoming law. As of today, the 2nd of August 2026, it is now the law of Europe. However, new companies have to obey now. Existing companies have four more months to not obey. But when they do, AI chat bots, including those support bots on people's websites, cannot secretly be AI. They can be AI. They just have to tell you, I am an AI bot. >> They can't say, "Hi, I'm Buford." They have to say, "I'm Buford, the AI bot." >> Exactly. >> Okay. Uh the same goes for basically online services publishing deep fake content. And this doesn't apply to stuff people post. This is about stuff or you know like media organizations post and stuff. If it's a deep fake, you have to label it as a deep fake. Basically, you have to label as AI generated. And there are explicit exemptions for art and satire. So you can be a satire website without having to put a disclaimer on every single obviously satirical video saying, "Hey, this is an obviously satirical video." And there is also none of this applies to generative AI you use to turn your best buddy into a Tyrannosaurus Rex, whatever you feel like doing, right? You can do all of that. Th this is much more about in Europe, you should be able to know when you're interacting with AI. It's not about stopping it. It's about saying, "Hey, oh, this is fake." Or not fake. This is generated. This >> generated >> and the explicit stuff. >> Go ahead. >> Go on. Okay. So, there's one more nice caveat in terms of text. If the AI wrote 99.9% of it, as long as there was human editorial oversight, it doesn't need to be labeled. It's only if you let the AI completely loose that you have to label it as >> not human. Basically, >> I think that's immensely sensible. So, they're not trying to stop us using the tools. >> It's just about a bit of honesty. >> I'm kind of trying to picture how an AI chatbot would do that. If I go to the Yuthi website and I have a question about a camera and I get a little text chat that I'm allowed to mess around in that first of all, I'm usually I like that it's an AI chatbot because a lot of times it has access like it knows the context of my question. It's not going to answer me for something from Wise or or some other company. It's going to be really targeted towards what I'm talking about. But how would that have human interaction in the answers? >> It wouldn't. So that's not where that's relevant. So that's relevant. If you're running a website where you have fired all of your staff >> and your articles are purely automatically generated, then your website has to make it clear that what this is AI slop. >> This is not human content. This is AI slop. The agent falls under the other category. They just have to say, "Hi, I'm Bob the AI agent." That's it. They're done. They're now covered. >> Good. Okay. A lot of Irish ones actually have a little robot icon which I think is probably arguably already compliant with the law. It's just like I am a little robot. It's like okay. Uh yeah. So I thought that was a nice development. A timely reminder given that it is the summer and some people are about to travel or some people are on what five out of six major trips done something like that. I lose count. If you're wondering that VPN, that reputable VPN I am actually paying for, not the free one because that's attacking you already because it's free. That reputable VPN I'm paying for, is it worth the bill? Here's a little bit of ane data just to remind you that yes is the answer to that question. There is currently an active campaign stealing Microsoft 365 accounts by being sneaky on hotel Wi-Fi. So VPN just yes, it's still a good idea. >> A reminder that AI is not all bad news because it's a bit easy given the conversation we had up here at the start of the show. It's a bit easy to get, you know, >> Google says AI helped Chrome fix 1,72 security bugs in their past two releases. That's better for all of us. That's amazing. >> That's fantastic. >> Yeah. And then we now get the good news bit. Lots of little things that are a little bit better. None of these are are chattering. Little bit better. Firefox and or Mozilla rather have had their quarterly big update. Firefox 153 has built-in containers for easy account isolation. So stop Facebook being logged in on every tag way more easily. So basically, you can use Facebook without it tracking you across the internet because Facebook is trapped in a little container and it's all nice and userfriendly instead of it being difficult to do. And Thunderbird 153 now has native support for Microsoft Exchange. So for everyone like me who detests the bloody Outlook app, you now have an opensource nice app that you can use for your Microsoft email. I think this is fantastic. >> Okay, >> Microsoft Entra ID. This is the identity provider for every corporate Office 365 user on planet Earth. They are moving as of the 1st of September pass key by default. You create a new account, it will default you to a pass key and make you choose something else. It will still let you, but the default the I'm not thinking. I'm just going to click next. Pass keys. >> Pass key. Oh, brilliant. >> I I have found one Pasi I really like. The App Store Connect. >> Oh, >> no. I still run into stumbles on that sometimes, but GitHub's number two, but the App Store Connect, when I go in there, there's a button that says Pasi. I click it and it takes my fingerprint and I'm in. It's beautiful. Nice. >> On the phone, it go asks me twice. Like the popup says, "Did you want to use your Pasi?" It's like, "Yeah, that's why I clicked Pasi." >> That's That's interesting. Okay, that's nice. I'm glad to see Apple, you know, I mean, they are one of the leaders of this whole pass key movement. They're big part of the phto alliance. >> Good. >> Yeah. >> Where was I with my good news? Proton >> release revealed that they got 47 data requests for VPN users and they said no to all 47. Zero data requests answered. So, we say it's a privacy protecting VPN. There's a receipt. doesn't guarantee they'll never say yes in the future, but it's a receipt. Um, if you will insist on carefully dipping your toe into this whole AI agentic thing, one of the things you inevitably have to do is let the agent authenticate as you to something. And right now that means giving the agent your password. That's very scary. One password and claude have worked together to give you a better alternative. You can allow one password to log Claude in without Claude ever seeing the details. So basically Claude hands over control to one password which logs them in and then hands control back to Claude. So the agent is logged in but they never had your details. >> Oh, that's interesting. Yeah, I clicked that link for sure. I want to read up on that. Yeah, great to see one password doing this kind of cutting edge stuff to still be relevant. Is it worth it over Apple's free passwords app? Well, for regular folk who would never have had a password manager, the passwords app is great, but there's still a lot of value from one password for advanced users. A lot of value. And this is just an example. Okay. Um, Google are making it a bit easier to get back into your Google account if you lose your authentication details. You can choose, and this is a 100% opt-in, you can choose to register a selfie video up front and then you can use a new video to get back into your account. So, they're using AI to prove you are you. >> Oh, >> which I think is nice. >> That's cool. Yeah. >> Yeah. And then another nice one is um Oh, why this ended up on the same bullet as the previous one. Um >> I'll fix that. >> Yeah, Facebook have added a new verified badge. It's not like the Xverified badge where it verifies that you pay money and not anything else. This isn't verifying that you're Bob. This is verifying that this account belongs to a human being. You're still completely as anonymous as you'd like to be, but the badge says this one's a human and they basically make you prove your humanness and then they give you the badge for free. >> Okay. >> So, I think that's great. You know, is this a troll account run by AI somewhere in >> I don't know, Russia. >> Perfect. >> No, it's a human. Great. So, it's nice. And finally, WhatsApp their web client. So, just taking a call with WhatsApp on the web. end to end encrypted. Now, >> excellent. >> All right, we can cleanse our [laughter and snorts] palace. Now, Allison, you gave me this first pallet cleanser a two and a half weeks ago, and I very diligently put it in my link manager and put it in the show notes, but I'm looking at the headline going, "How is this a cleansing my palette?" [laughter] >> Well, I don't know if it is, but I've been intrigued by the headlines. There have been a lot of headlines saying that big companies are all blowing smoke about them saving any money or anything with uh with AI that they're just, you know, telling their employees, you better use AI or you're fired. And and I've been reading a lot of these articles, but this is a a a site called the board.world. And this article is explains AI replacing jobs in 2026. 14 professions have already been eliminated. So, this isn't a happy joy joy story. I'm not quite sure why we ended up with impallet cleansers, but I want to talk about it. And they they give intricate detail on every one of these topics, but for example, 70,000 roles have been eliminated in text tax preparation because AI AI can do that, right? Entrylevel coders, GitHub copilot ate the junior dev pipeline. I don't think that's a surprise to anybody listening here today. parallegals and legal assess assistance 115,000 US jobs at risk goes on and on but it gets down to a couple that are real real interesting radiology the machine sees more than the doctor in this article they explain a 2024 Nature uh medicine story found that that the AI systems detected breast cancer and mammograms with 9.4% 4% lower false negative rates and 5% lower false positive rates than two rad a two radiologist consensus panel. So they're doing >> that's an interesting data set right because I can tell you there's another study that I was fascinated by. >> So the AI on its own is better than the humans on their own and that's what's in that study and that's been verified by quite a few studies. But you know what's better than both of those combinations? A review of one human and one AI. >> Interesting, huh? >> And that beats the AI zone. >> That would cut the number of radiologists in half. The reason one of this one was in particular interesting is the median compensation for the people whose jobs are at risk because AI is better at it is $480,000 per year. So these are not entry-level jobs. These are people who are highly trained, have gone through years and years and years of school, and two of them together is not as good as one AI. So anyway, I don't think it should have been pallet cleansers, but I thought it was a fascinating story. >> Well, I'll tell you what, Allison, for the sake of future readers, do you mind moving that link into interesting insights, keep the from Allison badge on it and just pop it up there in interesting insights because >> um I deleted that section, so I'll have to go back. Maybe I'll let you do that one because I don't know where the >> I'll move it. I'll move it and I'll put the bridge in. >> Okay. So, let's cleanse our pallets. Bart, I >> was going to say I'm gonna let Let me think about the right way to do this. I'm going to let you push your changes before I do anything or I'm going to make you a conflict that we don't want to have to deal with. >> Right. >> So, I'll let you do your bit and I'll fix it. And >> that won't help you. >> That's okay. I I >> Okay, we'll figure it out. >> We'll figure it out. >> Okay. So I have some I have some multimedia suggestions. So the first one is a podcast to listen to and this one may not be quite happy joy joy but it is fascinating and I consider a fascinating bit of learning to be a palette cleanser. So this is this was linked to me. Who linked it to me again? Um oh sugar oh no sorry I'm getting my stories confused. Pause what I just said for a moment. Um so the first one is a podcast. Um, and this is a purely happy happy joy joy one. It is Noble Blood is the podcast that hosted the episode, but they they do kind of an interesting crossover called Very Special Episodes where it's not actually part of the Noble Blood series and has nothing to do with that podcast, but it's the same host. The podcast is called How a Dog Talks: The Surprising Redemption of Comic Sands. It is a fascinating story talking to the actual people who invented Comicsands and it's a reminder that Comic Sands is an intentionally bad font. That's not a bug. That's a feature. So, the problem isn't people using Comic Sands. It's people using comic sands in context where an intentionally bad font is not appropriate, like a funeral announcement or a corporate memo. The font was designed to give a feeling of playfulness by being intentionally sloppy typography. >> Oh, interesting. >> So, all of these critiques from all of these high and mighty people about how the kerning is awful, that's the point. So, I thought it was fascinating and I thought it's a kind of a nice counterpoint to the reflexive everyone who uses comic sands is always an idiot. It's like no context, folks. Context. And it's really nice to hear an interview with the people who actually invented the font and what they were thinking and how Microsoft almost killed it cuz Microsoft management were like, "But the shapes, they don't match up. It's sloppy." And you the guy kind of go, "But that's the point." >> I love that. How long has that person wanted to tell that story? Or maybe it was funnier not not having anybody know all these years. >> And it also gives a tie back to Microsoft Bob. Believe it or not, the font was created to make the Doug in Microsoft Bob talk. [snorts] >> That is a real call back. >> Anyway, >> uh this one is fascinating but not quite as happy happy joy joy. So it was six colors. So Jason Snell pointed me at this one and it's available as a video podcast or a podcast podcast. Um it's the podcast is called Origin Story and they look at history of things to explain where stuff comes from and this episode is called Rage Against the Machine because it's about the Ludites. So we use the word Ludite completely wrong. We use the word ludite to mean someone who reflexively hates technology. The actual ludites were not against technology. The actual ludites were like the people who worry about AI taking away jobs. They were interested in workers welfare not in stopping technology. They weren't trying to stop technology. They were trying to stop technology ruining people's lives. >> They wanted to do it better. And so I am a proud ludite when it comes to AI. >> Agentic AI in particular. Yeah, that's great. [laughter] >> Let's start calling ourselves lites and see if anybody remembers this part. [laughter] >> Yeah, no is the answer to, oh, I won't bother listening to part. He's a lit. It's become such a porative, right? We we use it to insult people to end conversation. But the history of the ludites, because they were real, >> is very different. Now, one of these days you're going to tell me sabotage is not about throwing the sabot into the uh into that was true. Okay, that's true. And silhouette is genuinely the French nobleman who liked having portraits of his shadow done while people were starving and so they literally cut his head off actually with a guillotine and they called those kind of shadow portraits after him. >> That's true. >> Wow. Didn't know that one. All right, I have successfully recreated interesting insights. So, no action required. >> Yay. Excellent. I don't know what >> that was. A long one. That's a whole show. That's an hour and 18 minutes, Bart. >> Oh, wow. Okay. Well, I will stop talking and remind everyone just in case you've forgotten. Until next time, stay patched so you stay secure. Well, that's going to wind up what might be the longest security bits of all time and the rest of this show. Did you know you can email me at allisonpody.com anytime they like? Hey, you want a money on distribution for those travel logs? Send it right there. I hope you do it early because I it bugs me when people do it later and then I got to go back and give them the older ones cuz it's a lot of extra work. So, do it soon if you want in. If you also have a question or suggestion, just send it on over. I'll be bringing my laptop with me on the trip because that's where I have my fun. Remember, everything good starts with podfeed.com. You can follow me on Masttodon at podfeat.com/mastadon. If you want to actually see Steve and my podcast work on YouTube, you can go to podfey.com/youtube. If you want to join the conversation, you can join our Slack community at podfey.com/slack where you can talk to me and all the other lovely noilic castaways like Bart and Alistister and Joe from the Northwoods. You can support the show at podfey.com/patreon like Raleigh or with onetime donation at podfey.com/donate. You can use Apple Pay or any credit card there. Or if you like PayPal, [music] we've got that too. podfey.com/pal. And if you want to join in the fun of the live show, you're going to have to wait until [music] August 23rd to head on over to puffy.com/live on Sunday nights at 5:00 p. p.m. Pacific time to join the friendly and enthusiastic Nosil Castaways. [music] I'll miss you live show people, but we'll see you then. Thanks for listening and stay [music] subscribed.