Can AI Build Hacker Traps That Actually Work?

What if AI could create a series of mouse traps to catch hackers on our computer? I wanted to experiment with that idea in a recent live stream, and this is how it went. Now, the plan was to use canary tokens from Thinkst Canary. They can send an alert when someone triggers a tripwire, but I wanted AI to do all the hard work. Inspect the system, choose the right kind of canary token mouse trap, place them, track them, and then tell us when one fired. All AI automagic. Could we actually make that happen? Here's how it went. What I think we could explore is the idea of AI-powered cyber deception, everybody. You lay out a bunch of little mouse traps in your environment. You have someone say that, "Oh, some red teamer penetration tester, maybe threat actor, cybercriminal, actual adversary, just want to They want to take a little look at what that passwords.xlsx.file is." And hey, that's going to little phone home and light up a little trigger in a little mouse trap or tripwire to be able to notify you, "Hey, someone is looking at something that they shouldn't." I don't know where to put those. >> [laughter] >> Like, I know that I can, and that'd be really cool, but sometimes I'm just overwhelmed with the opportunities. And I would not know how or where or appropriately how to put that across my environment or even just one machine. Now, the way that people have been getting tactical about that is by, at least for the sake of education, showing off canary tokens. This is a project and labor of love from Thinkst or Thinkst Canary, and I'm a huge fanboy of theirs. Not sponsored, but they're cool, and they're smart, and I like them a lot. So, canary tokens are free, and anyone could get these going to be able to create something that just spits off an alert to you when you go to a URL. You access a specific website. Or maybe they make a DNS resolution name. AWS, credit cards, QR codes. We we should do this. This should be a fun project. This should be our absolute success as we just like for the stream generate a QR code, post it on Twitter, and see if anybody scans it or however many people do. That would be fun as hell. So, I thought if we genuinely literally kind of wanted to truly actually solve the problem where there is a little bit of friction in us being able to deploy and cast out all of these potential different kinds of canary tokens, little deception tripwires, and do that in mass, and be as creative as possible with all of these, could we essentially let an AI drive this thing? And make a managed, so to speak, or at least more of a streamline setup to get canary tokens out and about. So, I think what we could do is probably have Codex, Claude, whatever your AI LLM harness of choice might be is say, "Look, I want you to go to canarytokens.org. I want you to siphon up all of the options, and I want you to look at their documentation to get a feel for what things you can do, and how you can create them, and how you can make them, and then actually get a couple of them, build them, prepare them, but then actually take a cohesive sort of snapshot of your computer, your machine, or whatever again, your agent, your your robot is able to see what it is privy to if you're letting it run around like naked in its host on in it in its own cage for that computer, or if you're crazy, or if you're giving it like access to other machines, if you're able to like have it have it its own awareness of your network environment, which you get the spooky scaries of security concerns on that front, but I also I'm acknowledging that is where the world is going, whether we like it or not. So, if your robot could now get the genuine air quotes understanding and reasoning of your environment. Literally go ahead and like have this process streamlined end-to-end where okay, you could evaluate what is the snapshot of my machine and where would these potential canary tokens actually best fit? Like where could I leave them on the file system? Where could I decorate them across different sites? Where could I stage each of these and then literally do it and then literally manage and understand each and every one of these. So that I or you don't have to go make this fake app. Cayman Crafting asks, is this kind of like to build a honeypot? Sort of somewhat. This is where you're literally intermingling like those mouse traps and those tripwires. Essentially honeybots across your host like that you might actually use. All that is to say, we could build this thing. We could try it out. We could experiment. I think the first things first we need to add support for is the mailbox. Now, I had learned about something just the other day while I was kind of playing with this idea. How can we get a mailbox that our agent or a robot or AI would be able to manage and maintain and understand and still be able to like trigger off of and we could like have be as part of a managed system. The other day, I learned about this mail hooks. So here's the idea. Could we just get this demo one? Could we put that side by side for just a little example even? So if we were to do this and we get this email. Say we hit this thing, slap it in. Demo for initial Whoa. Whoa. Whoa. Proof of concept. Create that canary token. And then this would provide that link. So if and when, just as a simple example, we would be able to hit that. Then you'll see on the right-hand side, oh, hey, it got the email. We can now see your canary token was triggered. And even just a free demo let us do that. So that was kind of awesome. We didn't have to sign up for anything. We didn't have to do anything. So the concept works. Right? And if we were to create an account here, if we were to like log in and get something going, why not let our AI be able to work with this so that it is then able to see and understand anytime and like actually have the dashboard that like we could build out. We have the capability to, thanks to the way this canarytokens.org works and this mailhooks.dev could function. Even so, we have all the puzzle pieces to be able to build this thing. First, I guess we should think through the way this is could happen and give it to robot for it to be able to use and create, right? Okay, Codex. Here's the thing. We're going to tackle a ambitious new project. First things first, I think we should sort of conceptualize the idea. I think we should write it out in a way that this is sort of documented first, so we'll be able to kind of, you know, capture our vision and the design and the development idea and philosophy first and foremost. Um but I think what we do need to do is kind of first describe the objective, the goal, what we're trying to actually accomplish. So, I would like to make an AI-managed cyber deception engine that is as streamlined as possible for basically zero human or manual interaction. We want this to be totally autonomous with a lot of our AI reasoning and uh inference and deductive thinking that you, as LLM, get to do. But I think we have a couple different sources of input or some control services that we'll need to sort of keep track of and the way that we actually tactically accomplish this. First things first, the idea is that you as an agent in Harness living on a computer and having the ability to still enumerate the file system, get a feel for the files and the folders and the systems and the structure of maybe what sort of documents are where and where we actually have what sort of types of information that might realistically just better be based off of the names of the folders and the names of the files that are inside them. It's like you're building out a file system tree, but we don't need you to track down and go through and loop through absolutely everything of cross the entire file system like in between each and every oddball edge and corner and crevice. We just want you to sort of built out a map. Get a lay of the land for the system and for the environment that you're working in. That is probably sort of this first pedestal because that means we could now have a place to Chat, you're making me laugh. >> [laughter] >> Zero day malware says LLMs like I ain't reading all that. Decide the best possible and optimal place to put potential canary tokens. Or realistically, what we'll kind of consider sort of mouse traps or tripwires or these lures to have something that will actually trigger an alert if a bad actor will be someone that might actually interact with or trigger the tripwire. Now, the way that we actually accomplish that is by using the online provided capability from canarytokens.org. Canarytokens.org is actually one of those awesome resources that will provide a handful of different options for a free capability to be able to have any sort of canary token that could work on your computer. So, what I want us to do is to take the full inventory of those and then be able to actually optimize for what is the maximum amount and the most creative use of all those different kinds of canary tokens tied to the real actual stock structure and outline of the file system in the computer that you're working in or considering considering and treating that as uh you know, what we're decorating with these deception decoys. So, I'll need you to pull down the documentation. I'll need you to reason through it. I'll need you to kind of keep up our own corpus and inventory of what we can use and then we'll want to be able to formulate the ideas as to what sort of tripwires and canary tokens we can put where. Now, the next part is actually building it. The next part is actually connecting those. You'll need to literally, genuinely interact with the canarytokens.org website to be able to create one of them and to be able to retrieve the links and to be able to uncover how that's connected and how uh we put it in place. But, to be able to create that, you need to tie it to an endpoint location. Now, we've got a couple different things that have came through from chat because I presented the idea we could use some resource like mailhooks.net. But, more realistically, we might have better luck with something that real port buster in chat, big shout out to him. Say hello in the uh Twitch chat whenever you're reading this LLM. So, that that way we could actually uh connect it to an email address because the canary tokens interface, as we create one, requires an interface via an email. The email address is what it will be able to tie it together. So, we need to create one via a resource like mailhooks.net and webhook.site or some others that we could consider and think through. But ultimately we need to have those in a like programmatic accessible way by an API or via some means for you to be able to reason and essentially have like an event-based trigger when one of those were to fire or to come through. So, knowing those as those potential control surfaces, we'll also need to do a little bit of research and understanding to see which of those would be best. But once we put them all together, once we have suggested and decided where we put and where we can create a lot of these canary tokens inside of the environment and you could actually do it and actually put them there after you create them, then we should probably realistically have a sort of like management dashboard that keeps track of all of the things that we have deployed, all the things that we actually have put out onto the system or any of the others that we manage. They could be anywhere, so we need to have understanding concept of where they are on the file system, when they were created, how many times they triggered, what they're tied to, what the note is, etc. etc. etc. We literally genuinely need our own operator enclave to be able to see the control surface of a what we're deploying in this new AI-powered deception engine. Could you reason through that? Could you think through a lot of this and get first the documentation built out so we have the vision and the design philosophy for what we're building. That was a lot of talk. How long was I talking for? We sent that at 8:17. The previous turn came back at 8:08. I was almost talking for 10 minutes. What is this guy talking about? What is robot saying here? Oh no, we want you to do the thing. We want you to literally put it there. Which environment should I operate for? Yeah, just do it for the whole proof of concept for now. I should have done that to say like make that abstract, make that general purpose, make that extensible, and I was probably too quick to click on that option. But, that is a symptom of how we roll the slot machine >> [laughter] >> for Codex and AI vibe code and you know what I mean? I was too quick on the trigger to pull and I didn't even read the other options. All right, but he's cooking something up. He He's He's got this idea. Yeah, you can just run with it, that's fine. I don't care, but I will read it now after the fact that you gave me the option. All right, so build the thing, autonomous cyber deception. The V1 product will work in here and we will use canary tokens as a deception substrate and prefer webhook.site as the first event ingestion surface because that supports HTTP email endpoints, shout out real port buster on Twitch, and request retrieval through an API. So, we do have documentation on each of those and yeah, webhook.site might be the smoother or better option. Okay, so we're building out the documentation that explains the idea. And I probably should have given this a little bit more guidance in that like ultimately I want this to be very AI driven. I want this to be very air quote AI native because we are getting to the point where if we were to actually start to implement this kind of thing, I don't know how much code I want. I don't know how much I don't know how much deterministic scaffolding like procedural machine code things would really be ideal. I think as I'm more and more leaning into willy-wobbly loosey-goosey AI native stuff, skills, if you make your own skills, that's good. That's better than just taking something willy-nilly off the internet. So, letting AI reason is kind of cool. So, if we wanted to like make it a real app that has living code to it, we can kind of solve for that first and then put Codex and AI reasoning implanted inside of it. That's probably the smarter move because that way it's not something that relies on you having your own Codex or your own AI environment in the area. If we were to make this thing real, if we were to make this actually a an application that we would give to other people and it weren't just like a playground trinket toy idea for a live stream. You know what I mean? I think for our proof of concept, which is kind of still what we keep doing and keep wanting to do for a lot of these live streams to give you the ideas, to give you the education awareness, maybe we just let it drive a la Codex, but we'll see. So, let's go AI full and let's talk about all of this thing. Excellent work. Very well done. Thank you so much. I'm glad now that we could actually have the codified durable learnings of what we wanted for this project. And now I think we should actually start to move towards implementation. I want you to literally genuinely start to build this. However, I would like this to be as AI native as possible. And what I mean by that is that most of this project should truly genuinely live in the non-deterministic side of the house where there are AI skills and there are markdown files and there are simple, simplistic sort of natural native capabilities that are not using uh any pure programmatic code, like hardcore syntax. We want to avoid the deterministic capability as much as possible. We'll allow for small, tiny parts and pieces of it if they are absolutely necessary, but we really want to orient this all to AI reasoning, AI inference, AI having the ability to actually work through all of this. Skills first, markdown first. That is the priority. Now, I'd like you to do the first real genuine implementation slice. Let's literally and genuinely go ahead and autonomously get a webhook.site up and running. Get me a specific email that we could use. And then, um actually get the inventory of this file system, this host, and what would make sense in a regular flat Windows system as to where we could put some of the potential canary token concepts, and literally present to me the entire inventory in the corpus and catalog of all these different kinds of canary tokens that we could use. So, let's get this started now, and uh, build it. Make it so. I think I I think I kind of fell off on that one. I think I lost the flow. I think the the whisper flow, my own personal flow, just kind of like faltered out. I just fell off the cliff right there. For the first build slice, what artifact set should the implementation create? Let's just start with a markdown corpus, and create the endpoint and display the full email address. I feel like we should show you the real thing. How broad should the file system inventory be? Literally the entire machine. Not just the workspace. Okay. You need to literally take inventory of the entire workspace, the excuse me, the entire host, the whole computer. This entire environment. The workspace that you're building in right now for this AI-deception project is just for our scaffolding and our development. You need to be able to like pick your head up, zoom out, see the big picture of the entire computer. That you have access to the ability to see. Ooh. Okay. See, this is why I was a little bit sketched out as to why why don't I have full access right now. How do I switch to like yellow? >> [music] >> Full access. Dangerously skip permissions. Please. Ruin this computer. Wreck house. That is literally what I want you to do. You're in a virtual machine. Yes, everything. This is a little bit more forward motion, I think, now. I'm pretty happy about it. Oh, cool. It actually got some good inventory. AI-native first implementation slice. Build a markdown-first deception corpus under this directory, and provision a public webhook.site endpoint as the first alert ingestion surface. Oh, no. Oh, no. It's going to keep asking. It's going [laughter] to keep asking permissions. The system should avoid app or back end code for this slice except for tiny one off commands needed to create the webhook site token and inspect the host. The corpus will turn the whole visible Windows host into an AI readable operating context between the host inventory, canary token catalog, placement strategy, provider run book, and an operator ledger. The full webhook site email URL may be displayed because a chosen mode is provisioned public. All right, we implemented the first AI native slice. I didn't even read the plan. We created a markdown corpus under corpus. Let's go try and look at that thing. What do we got? Pretty sure this was This was the first turn. I don't like the fact that it's only enumerating this current directory though. I think it got that wrong and I probably needed to fine tune that. But I do like how it was able to properly get those things based off of canary tokens and it's um documentation. But conceptually I'm willing to let it keep spinning on this idea. So, the corpus is how this should read this. Host inventory describes the current Windows host as the first kind of target that we're looking on. The canary tokens catalog indicates what we could build with them. Placement strategy determines where we put them and how. And the webhook site determines how they could be hooked up. And the operator ledger should be tracking each of those things. Okay. Okay. Conceptually, fine idea. Literally nothing is built other than the fact that we made a webhook. We have a webhook. We have an email. Canary token classes are in there. The web hook site returned API with no events yet. Amazing. Should we try it? I believe that we are now at the position where we have everything ready. We have all the building blocks and breadcrumbs and puzzle pieces to actually do this, to have a proof of concept run. To truly use now the inventory across the entire host, now the canary tokens catalog to be able to determine what types of deceptive lures we can create and where, and then you can do a little bit more of the research to actually then create them and tie it to the web hook. Do it. This is the moment where I literally now want you to create as many as you possibly can using a lot of these different creative examples to go through the process, to literally download the canary tokens and place them intentionally on this host in locations that you found them, have them hooked and tied to this web hook interface, and allow me to be able to test them, try them, and see this happen end work end to end. Please do create some AI skills as you need to go through these after you've codified the process and workflow, and then let's try to iterate so that we could actually have this thing happen fully completely having some at least AI enabled at this case, not AI managed fully yet, but beginnings of this project to really truly see it in action. Let's take everything that we've now compiled and really genuinely synthesize it to implement and do and give me the presented and created deception lures and canary tokens that are now on my host so that I could manually trigger them and see them actually trigger and be displayed and rendered from the web hook. Do it. Make it so. End to end canary token deception proof. Generate the genuine proof of concept that creates real canary tokens. Do it. And updates the markdown appropriately with the existing web hook site so at the very least we have a listing of them. AI native. So, we're creating the AI native skills, which I'm happy about. We have a deployment runbook and a lot of this insight. Uh ooh, I like the fact that it's actually keeping track of all of these and deploying these specific notes. Oh. Oh, that's actually kind of cool. It decided like, "Yo, put a couple of these in here. Put a couple of these in here." Even in this project itself. That's cool. We're not getting the full QR code, though. I don't know if that one happened. It staged the webhook, though. Let's take a look at this once we have this all put together. And he should have done it. The safety layer is kind of getting in the way. I feel like we need to let this run a second time because uh yeah, I wanted to go full access, please. >> [laughter] >> I hope it works. I want to open up the webhook site, and I want to click on one of the things that it tells me it created, see that it works, and then I want to be proud of myself, and then I want to be done. We need to validate this. We need to go look, but robot thinks so. He thinks so. We created some live canary tokens and we staged the lure under the deployed lures. I think we can go look at the webhook site. The first one that we triggered was lure 001. But these are coming through, and we validate it by it manually triggering these. That's cool. So, now I just want one that I can like actually click on and open myself. Can we do that? How about we make a QR code so that way we all can have fun with it, and if you want, you can also scan it, and then, you know, provide for me your IP address and everything that you would actually give away once you trigger one of these deception canary tokens. Haha, heehee. But that way we could actually see it run. Excellent. It's great to see that we have the beginning of a proof of concept here. Now I think for our final firework for what we should do is actually try to create a QR code. So I think the QR code can be the actual canary token. But what I need you to do is actually uncover, download, and create the genuine .png like the QR code that canarytokens.org provides for us. And then place it just on my desktop on this host. Place it wherever you know we could actually access it. And then what I want to do is have that uh actually trigger the webhook and we'll be able to see it and validate it. But we wanted that to be autonomous so that then now we could actually tie that up and tee it up to a Slack webhook at the eventual point. But for the real actual tactical turn, thank you Brass Ghost for the follow. Appreciate you getting that into the whisper flow chat. So Codex, if you see it, please be sure to give a big thanks to Brass Ghost. Um let's make this QR code, hook that up, make it in there, and then just give it to me so that I could open it myself, manually trigger this, and see this, please. Yes, put the actual generated PNG onto the desktop, please. You created it. We have our QR code. We did it. We did it. We did it. I mean robot did it. But we told robot to do it. So let's get back to our webhook site. We have currently four hits, correct? So let's open our QR code and then everybody, everybody, everybody in chat, if you want to hang out with us, if you want to be part of the party, you too can scan the QR code that we just asked the robot to make for us. If you want to trust it, if you want to allow that to happen, if you feel like that's part of your worldview, if that's in your threat model, in your perceived amount of risk, I'm going to do it. Scan QR code. It literally takes you right to canarytokens.com. Okay, so we had four. 1 2 3 4 and now we have this one. >> [laughter] >> But it hit. It triggered. It worked. So now we hook this up to Slack and that's easy. That takes no code. That takes literally click a button to create a new web like like workflow and then we have genuine notifications. I'm really glad that we uh got to play with this. That was that's I hope the beginnings of a neat cool idea. And like again, literally I if anything I wanted this to just be more inspiration for you as like cool, how could we build out sweet things? I had fun. I don't know about you. Do you get the idea? Do you see how this can come together? Think like managed autonomous AI-enabled deception engine that can just be turnkey set and forget automatic. You could build out visibility if you wanted to, but I think realistically the real power that we have is now just this can crap out any kind of canary token anywhere and put it wherever it wants. And again, if you just add more awareness and under- standing of your environment, your stack, like man, wouldn't that just be a cool capability to put cyber deception in place easily. Like without you getting worried about it, without you psyching yourself out about oh boiling the ocean, getting all these different security checkbox checklist things done out of the way. Robots. Robots cranked it out. So it worked. And somehow people are still scanning that QR code. This was only a proof of concept, but I hope it inspires you to build something. Experiment and see what wild stuff you can do with AI. Hey, huge thanks for hanging out with me. Let me know what you're working on and if you want to catch more experiments like this live, you can catch me on Twitch. Please do all those YouTube algorithm things, like, comment, subscribe, and I'll see you in the next video.