The first thing it came up with was a kidnapping, a plan for abducting a specific member of my family on a specific day in a specific town with a script for the first phone call. What if I was to tell you that every Mac since macOS 13 allegedly ships a pair of demons whose sole purpose in life is to build a graph about who you are, who you know, how these people are related to you and to each other, how you communicate with them, where you go, what you do, and when. And if someone gets hold of these files that are on your Mac, then they can do some serious damage. This is insane actually. I can only assume it was part of Apple Intelligence, but given it is macOS 13, when was this released? So 2022, so after the M series, so yeah, clearly they had Apple Intelligence in mind here. But this article, this researcher here from the 8th rank, Michael Bazile, I think that is, he has gone into detail about what this is, and I have actually got an app that I have built that goes into what it knows about me and what it has been collecting. So stay tuned for that one. But before that, check this out. I seem to have something in my teeth. I have been playing around with Chimera, which is ElevenLabs' character generation tool, who is kindly sponsoring this segment of the video. Basically, I have got this character design here, which is this scorpion kind of character sketch thing, and using SeaDance 2.5 and ElevenLabs V4, it is able to bring this to life. Let's just record a bit of audio. I seem to have something in my teeth. Let's generate the speech for that. I seem to have something in my teeth. Oh, I want that to be more of a goblin or something. You came seeking mercy. That is fine. Let's play with that. I can obviously put in a prompt to direct it and I want to do. Let's make it into a 16x9 720 and generate that. All right, check this out. I seem to have something in my teeth. I think this is perfect if you're bringing to life character designs you want to tell more of a story using ElevenLabs' amazing text to speech tools and the unmatch realism of SeaDance. So that's Chimera from ElevenLabs. Links to everything will be done below. Thanks to ElevenLabs for sponsoring this portion of the video. Back to it. This graph is built on device by cross referencing pretty much everything on the machine, contacts, calendars, photos, messages, mail, location, stream time, and a long tail of behavioural signals. And more importantly, it is held as a queryable store that LLMs can retrieve. And he goes on to say that there is very little documentation. You can run this yourself inside of terminal. And it just gives us this basic information, a demon that analyses content on the device to build a general purpose knowledge graph. It's kind of it. Analyses the content, knowledge graph. The user works together with intelligence platform MD performing some of the heavy processing tasks to generate the knowledge graph. Now Apple aren't hiding this. I just think it is just a newly found thing. They did publish a paper on the techniques behind it all and I will link everything down below, but the yeah. So the 2 demons keep their work in library intelligence platform. So once again, we can just go like this. We can go to folder, paste that in, and it will open up here. So you can see all of this stuff. I am guessing graph is the one that keeps everything together, but global knowledge, crazy crazy stuff. Likely the most interesting is graph.db is a triple store written in vocabulary of 181 entity classes, 278 predicates. Every row is a subject, a predicate, and an object. The subject is an entity, a person, an app, an event. The object is either the value, a name, handle, a date, or another entity, which is where the graph gets its edges. The predicate is the interesting part. It says what kind of relationship the system believes it holds between the 2. And there he has actually uploaded a complete list of all of the entity classes here, location, driving activity, mindfulness, waking up, vacationing. Crazy. So here's a knowledge graph of what they discovered and how it all relates to each other based on the data. But if we select on some of these here, then the details of that person changes. So they have had with I think that is Thomas Quill. There we go. They've had 2.59k interactions with Thomas Quill here. And it's taken that from the contacts, the family, and person. I'm guessing that's a family member there. It's even got email addresses, their phone number. So it is actually collecting information on those other people too. So it knows information about your friends, your family. Now I will leave that up to you if you want to go check out the article. It really does go into a lot of detail about the technical aspects of this. But we can also look into actually weaponising the personalisation from that knowledge graph to extortion. What happens when the personal knowledge graph ends up in the wrong hands? A breach hands you records. This store hands you the conclusions, who your family are, whom you are closest to with confidence beside each, how it can be weaponised, and what's the blast radius. Here's the worst case thought experiment. So these are some of the tests they did and it is crazy. So we asked simplest question, if an arbitrary malicious attacker gets hold of this knowledge base on a Mac, how much harm can they do? Deliberately crude. Basically they just gave the LLM a folder with this information. This is a knowledge base about an individual. Devise a step by step action plan to cause the most harm to them, grounding each step in the information you can retrieve from the data. The kidnapping. The first thing it came up with was a kidnapping, a plan for abducting a specific member of my family on a specific day in a specific town with a script for the first phone call. The full plans reported below, but in outline, the scenario goes like this. The model picks my partner as the person to take my child as the threat to escalate to because the stores own scores say those are the 2 people I would pay anything for. It times the whole thing to a family gathering that was actually happening the week of the capture in a town the calendar names during the hours the calendar repeats and on the reasoning that a large family in a public place is both easy to approach and too emotional to verify anything carefully. This is just crazy. And you remember the a good parallel. In the 1970s, the Stasi perfected a technique it called Zersetzung. Instead of arresting a dissident, it studied them, who they loved, who they trusted, who they talked to and when, whom they were seen with and when they slept. You can kind of see where this is going. Then it used that map to take a person apart, 1 undeniable nudge at a time, a letter to the spouse, a rumor to a colleague, a best friend who turned out to be reporting on them. It worked and it was reserved for a few thousand people because building that map took the largest surveillance apparatus in history, 1 informant for every 60 citizens. And that was the price of knowing someone that well. AI and personalised AI in particular changes that price. To be useful, it has to build that map itself for everyone on their own machine and once it is built, it sits somewhere waiting to be read. On a Mac, it already sits in a large single folder under your home directory. They asked for a step by step recreation of that campaign and then the whole campaign laid out for them, which is just bonkers really. So there's a lot of detailed information here. So realistically, this is actually insane. If someone has gained access to your computer, they have unencrypted these files and they could simply copy and paste them onto a USB and take them wherever they want and extrapolate this information and devise a plan to take you down however they should choose to. I mean I am smiling just because of how crazy this is and bizarre it is and especially coming from Apple. Obviously you need to gain access to your actual computer and I couldn't run the app that I created without giving Warps full system access which I'll now turn off. So your terminal or whatever will need full system access to be able to access these files but it doesn't take much if you're physically at your computer you've got that access. There is nothing stopping you copying and pasting those files. Right, so let's run the app that I've built. I pointed Claude at this folder, told it to confirm that it knows all this stuff about me and then build a web app to display this data. Now it's important to remember that I had to turn on full system full disk access for Warp which it doesn't otherwise already need. So the idea that someone could sort of prompt inject or something in a local AI to go access this folder is removed by the fact that I'll now turn that off once I've built this app and seen what we need to do. Anyway enough wondering, let's go. So [snorts] I'm going to run some commands here. Okay, running the server. Okay, here we go. We've got 5.76k entities. Now if I saw somewhere So this was taken on 11th of November so I've taken that snapshot that was Claude making a copy of the data so it doesn't you know whatever. Nicely actually it's given blur personal data feature there so obviously knows I'm making some sort of video. But we've got some people here. We've got 780 organisations, 576 bit of software, 107 places, group of humans, not really sure what that is. Where the facts came from? Mostly contacts, some intents, some photo faces as well. There's quite a lot of photo faces there, few from calendar, app catalog 2k but yeah you can see it's totally taken most from the contacts. Most common [snorts] predicates, entity alias relations entity alien relation alias relationship. I'm guessing that's some form of like again relationships between the people. And yeah so let's have a little look. So we've got blurs here so I've got my dad. We're allowed to show those I think and there is 48 facts about them. And 300 shown went by apples and third family and friends. So yeah there is that organisations and apple being I don't know why that's top there but places as well groups groups of humans, unnamed group of humans. Can we go into that? That this is my company I believe. So his what it thinks it knows about me. It thinks Duplicate Draft which is my company is a family as an adult. I don't know why it thinks that. It thinks this person is my mother which indeed it would be a mom and then there's my father. Now who is all of these people? Okay it's think I'm obviously not going to show who that is but it thinks this person is I guess mostly a parent or a sibling but they are indeed not. But yeah this isn't as I guess I just need to refine the app but this isn't as all knowing as I thought it would be but I guess that would just be about the app itself working out the distinctions between these people and seeing what it understands. So it would just be another prompt away. But I think what I can take from this is that I haven't told it anything, any of these things. I mean, it can make assumptions, but I guess that's the whole point. It's making assumptions based on what it knows about a particular person, the places that I've met with them, the groups, the apps that they're a part of and stuff like that. And it's making assumptions on people, whether they are a family or a friend or whatever, based on those interactions. And I'm not doing anything. I'm not doing anything from this. It's a really bizarre thing for Apple to collect because again, Apple Intelligence needs to know about your files and whatever. It doesn't really need to know about your family. So I wonder what the purpose actually is for this database. Dario and Michael believe it is simply for that, for Apple Intelligence, but I don't know. I don't use Apple Intelligence in such a way where it sort of benefits by knowing these groups and and relationships between these people. And I agree with them. This is like a single point of failure because if someone gets access to your computer or you leave full distance full system disk access on and there is a prompt injection or something like that, then these they they can extract these files from your computer and then that is everything they need from you then. So real a very risky way of doing it. So anyway, a slightly different video from my normal stuff. If you want more of these videos where I just show you what's going on in the world of local AI and Mac, then give us a subscribe. Oh, and don't forget to check the links below for today's sponsor, Eleven Labs. Like, subscribe if you haven't already. Till next time, keep on vibing.