Latest / Stage Talks with Bellingcat / "Using AI In Investigations - To Uncover A North Korean Hiring Scam" with Craig Silverman
Transcript
- 0:04You're listening to a stage talk titled using
- 0:07AI in investigations to uncover a North Korean
- 0:10hiring scam. This week we were joined by investigative
- 0:13journalist Craig Silverman. Craig took us inside
- 0:16his investigation of a sophisticated North Korean
- 0:19hiring scam from fake recruiters and job interviews
- 0:22to malware, online identities and cryptocurrency.
- 0:27And through that example demonstrated how AI
- 0:29can help investigators uncover connections. analyse
- 0:33evidence and work faster across vast amounts
- 0:36of online information. Despite his successes
- 0:39using AI, we discuss where human judgment and
- 0:42verifications still remain essential. You can
- 0:46find links to all the resources mentioned in
- 0:48the talk in the podcast description. This talk
- 0:50was hosted by me, Charlotte Maher, on Thursday the
- 0:531st of October, 2026 in the Bellingcat Discord
- 0:56server. Happy Thursday, all. Thanks for joining
- 1:09this week's State Talk. Today, we're joined by
- 1:12the wonderful investigative journalist, OSINT
- 1:15expert trainer and Indicator co -founder Craig
- 1:18Silverman. At the moment, it feels like the world
- 1:21is being pushed towards using AI, the next step
- 1:25in innovation across almost every industry. The
- 1:28media and research sectors are no exception.
- 1:31But is AI really a useful tool for open source
- 1:34research? Or can it become a time consuming hindrance,
- 1:38leading to more confusion in our news and fact
- 1:40-gathering rather than greater clarity. Today,
- 1:44we want to explore that question with Craig who
- 1:46has been putting AI to the test in real investigations.
- 1:49He's been using AI as part of his investigative
- 1:51work, including in his recent investigation uncovering
- 1:55a North Korean hiring scam. The link to which
- 1:58is in the chat and will be in the description
- 2:00of this podcast if you're listening back. So
- 2:03what can AI actually bring to investigative research
- 2:06and where does it fall short? We're going to
- 2:08talk about why AI could be useful for investigative
- 2:11research, where it may generally change the way
- 2:14researchers work and where we still need to be
- 2:16very careful about relying on it. As always,
- 2:19we'll have plenty of time to answer your questions.
- 2:22You can place your questions in the chat accessible
- 2:24on the right hand corner of your screen. But
- 2:27please remember that this is being audio recorded
- 2:30for the podcast. So if you don't want me to mention
- 2:33your username, please add that as a note in your
- 2:36question. Okay, Craig, I'll pass the mic over
- 2:40to you. Tell us about your work. Thank you so
- 2:44much. Hi, everybody. I realize people often start
- 2:47these things by saying they're really happy to
- 2:49be here and how much they love Bellingcat. But
- 2:51I am really happy to be here and I really admire
- 2:54Bellingcat. And it's crazy because more than
- 2:57a decade ago, I remember Eliot, the founder
- 2:59of Bellingcat and I, being in Amman, Jordan,
- 3:01giving trainings together. And so it's pretty
- 3:04cool to come here and be part of this great community.
- 3:07So thank you for having me. Uh, so look, the
- 3:10plan is I do have, um, you know, uh, a talk and
- 3:14some, some visual aids to go with it, but we'll
- 3:16make sure for folks who are listening audio only
- 3:19that, um, you get the full experience as well.
- 3:21And then I'm super excited to take your questions.
- 3:24And I will just say off the top. I mean, I'm
- 3:27not giving this, uh, talk as somebody who has
- 3:29perfectly figured out everything about AI. And
- 3:31I'm certainly not giving this talk as somebody
- 3:33who adores everything about AI and LLMs and the
- 3:37products and the companies behind them. One of
- 3:39my books was used as training material without
- 3:43my knowledge or permission. And yet here I am
- 3:47talking to you about how I'm using some of these
- 3:50tools to assist with investigations. And I think
- 3:52that is the key word here is assist. I don't
- 3:55hand things over completely, but I definitely
- 3:57am seeing some benefits in terms of, you know,
- 4:00saving time in terms of being able to run things
- 4:03on an ongoing basis and be able to especially
- 4:05do some more technical stuff that I personally
- 4:08don't necessarily have the skills for. So let
- 4:11me dive in. I'm going to share my screen and
- 4:14walk you folks through some things here. And
- 4:18as you can see, there's a link that I ended up
- 4:21publishing and that AI really helped me with
- 4:24some of my research process. So, you know, this
- 4:28is the story was published in June. Um, and it's
- 4:32one that I, I spent, uh, I guess, you know, several
- 4:35weeks working on, um, it wasn't a hundred percent
- 4:39of my time during that because at Indicator,
- 4:41we kind of do, uh, three things. One is, you
- 4:44know, we publish investigations. It's, um, you
- 4:47know, we're an independent outlet, two, I'd
- 4:49give a lot of workshops and publish a lot of
- 4:51guides to digital investigative techniques for
- 4:54paid members. And then three, we also do, um,
- 4:58publish and build tools. And I'm going to show
- 5:00you. a couple of those tools that we've built,
- 5:03a couple of which there are sort of free access
- 5:06on things for. And if you don't know what Indicator is,
- 5:09we're about a year and a half old, and it was
- 5:10founded by me and Alexios Mantzarlis, we both
- 5:13come from a background in doing fact -checking,
- 5:18digital investigations, investigative reporting,
- 5:20and that kind of stuff. So this story came out
- 5:23in June. Let's talk about sort of how I found
- 5:27it and how I started. And so the first thing
- 5:30was this, this started with a thread on X from
- 5:33an account where he said, I got completely owned
- 5:36by the most sophisticated hack I've ever encountered.
- 5:39He said, I'm a developer. I know what scams look
- 5:41like. This didn't look like one. Um, and so one
- 5:44of the interesting things that kind of caught
- 5:45my attention about this is it wasn't just somebody
- 5:48sent him a phishing link and he clicked on it
- 5:50and, you know, mistook it for a legitimate one.
- 5:52And then boom, they, you know, were on his computer
- 5:55in this case. he actually had in -person interviews.
- 5:59And when I sort of read this thread, it did seem
- 6:02to me like this was a classic kind of North Korean
- 6:04hiring scam, and yet it had some elements that
- 6:08seemed a little bit interesting and potentially
- 6:10different to me. And so I was interested in this
- 6:13and I wanted to learn more, but my assumption
- 6:15was that, oh, you know, probably the website
- 6:17is down, and yet in this thread, the guy actually
- 6:21linked to the GitHub. of this, you know, allegedly
- 6:24fake company that had tried to hire him. And
- 6:27on the GitHub, I had, you know, the great starting
- 6:30material for those of us who were doing, you
- 6:32know, open source work. I had not only the company's
- 6:36GitHub, but its website. It had an X account.
- 6:39It had a LinkedIn account. There was an email.
- 6:41When I looked on LinkedIn, there were several
- 6:44executive account profiles still there. The website
- 6:47was still up. And so at that point I said, you
- 6:50know, it's almost like I started a stopwatch
- 6:52ticking at that point because I said, all right,
- 6:54let me see what I can find and let me see if
- 6:55this is promising and worth me actually diving
- 6:58into. I didn't know if this was going to be a
- 7:01story I pursued or not at this point, but it
- 7:03had enough of this kind of raw material, open
- 7:05source material for me to at least be able to
- 7:08do some gathering and looking into that. And
- 7:11so this is actually where I first brought AI
- 7:14in. I don't endorse any particular company or
- 7:18model or anything like that, I personally do
- 7:21pay for Claude. And I typically interface with
- 7:24it through the command line interface through
- 7:26Claude code. Although I also use the app and
- 7:29co work. But everything I'm going to show you
- 7:32today was done really with me working in Claude
- 7:35code, which is in the terminal. And that's so
- 7:38what you'll, uh, you see with some of the visual
- 7:41things is, you know, the prompts, the requests
- 7:43I'm making are in the terminal, not in the sort
- 7:46of more user friendly app. And then the other
- 7:48thing that I'll note is that I did, you know,
- 7:52just a month or two before starting this particular
- 7:55investigation, I did actually shift where I used
- 7:58to use Evernote for my note taking and clipping,
- 8:00and I shifted to Obsidian. Maybe some of you
- 8:03are familiar with Obsidian. It is a free, very
- 8:06powerful note taking, but also kind of network
- 8:09analysis app that you can use. And so I had sort
- 8:13of done some work, which I'll show you, to kind
- 8:15of create a setup where Claude had access to a
- 8:19vault of notes and information in Obsidian. And
- 8:22so I was able to basically give Claude a whole
- 8:25bunch of context, literally about 15 years of
- 8:28my research and reporting, all of my thousands
- 8:31of notes that I had imported from Evernote. And
- 8:35this was all material that I was okay sharing
- 8:37with this tool, right? Anything super sensitive,
- 8:40I don't put in Claude. I'll say that as a kind
- 8:43of caveat right off the top. because it's an
- 8:46actual question. If I'm dealing with sensitive
- 8:48sources, sensitive documents, I do not give that
- 8:51access to Claude via Obsidian. I do not throw
- 8:54it in there. But in this case, I had loaded a
- 8:57huge amount of material into an Obsidian vault
- 8:59that I connected Claude code to. Um, and so the
- 9:03first thing I did was I downloaded this guy's
- 9:07tweet thread. Um, and, uh, as a PDF, I gave that
- 9:11to Claude and I said, Hey, I'm looking into this.
- 9:14Please, um, open an investigation folder. So
- 9:17I've created a sort of shortcut where, uh, I
- 9:20say, Hey, I'm starting a new investigation and
- 9:22Claude knows to open a bunch of folders for like,
- 9:25you know, names of people and entities and other
- 9:27elements. And I said, basically. scrape all the
- 9:30information out of this guy's thread, set up
- 9:33a note with everything in it in his name, and
- 9:36then start backgrounding the guy. And so that's
- 9:38the first thing that I did with using AI was
- 9:40I actually asked it to do a quick sort of poll
- 9:42of like, who is this guy who sent this thread?
- 9:45Well, I was following some of the threads of
- 9:48the websites and the other things like that.
- 9:51And so this is where I'll sort of tell you a
- 9:53little bit about this setup a bit more. And I
- 9:57don't... you know, preach this as the perfect
- 9:59setup. It's been working for me. I've actually
- 10:02started tweaking it because we've actually just
- 10:04launched literally today a bunch of tools for
- 10:07AI to help with investigations. But this setup
- 10:10is the one that I started with and have been
- 10:12evolving from. And one of the key things that
- 10:15I wanted to try to address with this is I have
- 10:18been using LLMs already and where I would go
- 10:21and I would make a request or whatever. But what
- 10:23I wanted was A kind of system that was rooted
- 10:26in my work and knew things about me in the context
- 10:30of the work that I do. And, and, you know, the
- 10:33publication I'm running so that I didn't have
- 10:35to start from zero every time, you know, and
- 10:37this is one of the realities as investigators
- 10:39is like, we have a history of work. We have methodologies,
- 10:42we have tools we like to use. We might have tools
- 10:45that we pay for as well as free ones. And to
- 10:49sort of work with an LLM and not have it know
- 10:51all of that from the start is, I think, a real
- 10:53disadvantage. And so what I wanted to do was
- 10:56to have a system that worked where, in this case,
- 10:59Claude could actually not start from zero every
- 11:02time, but have the context. And I also want to
- 11:04get past one of the big problems with these systems
- 11:07is the sycophancy, which drives me crazy. Um,
- 11:11where, uh, you know, I will have it go and pull
- 11:14some information and it will come back and it
- 11:16will sort of share what it pulled while adding
- 11:19that into an obsidian vault, which I'll show
- 11:21you in a second. Um, and at the same time it
- 11:24would come back and it would say things like,
- 11:25Oh, rich findings. Oh, this is very promising.
- 11:27And I don't, I don't want an LLM to do the analysis
- 11:31work. And I certainly don't want it to like be
- 11:33glazing me all the time about what a great and
- 11:36genius person I am. Um, and I certainly don't
- 11:39want hallucinations. And so, um, what I wanted
- 11:41to do was create a system where I could give
- 11:43it instructions and sort of, you know, have it
- 11:45understand things about me, but also how I want
- 11:47to work. And so I have kind of a diagram, but
- 11:51at the core of this is what I've kind of already
- 11:52described, which I is I have a vault. that sits
- 11:56in Obsidian, where I have about 15 years of articles
- 12:01I've written, which I imported, typically as
- 12:04XML files. So, you know, I used to work at ProPublica.
- 12:08I worked at BuzzFeed News before that. I spent,
- 12:10you know, a long time as a freelancer and contractor
- 12:13writing for different places. And I pulled in
- 12:16that archive of things that I'd already published.
- 12:18And so it's there. And again, my system, Claude,
- 12:22has seen the work that I've done. also has notes
- 12:25and non -sensitive material from those articles
- 12:29and things that I had gathered. And so all of
- 12:31that context is there. But two really important
- 12:34files that I'll show you in a moment are a clod
- 12:36.md file and a haunt .md file. These are markdown
- 12:39files, which these AI systems really parse very
- 12:43easily. And so I have all of that history and
- 12:46context. to tell her how I want to work and what
- 12:48we've been working on and the kind of work I
- 12:50do and who I am. But then also the stories and
- 12:53investigations, the nonsense of elements of them,
- 12:55they also sit in obsidian and Claude can see
- 12:58all of that. And in fact, can not only read it,
- 13:00but write into it as well. And so I work in the
- 13:04vault and my AI agent Claude Code can also work
- 13:08in the vault. And then I have also as a backup,
- 13:11I have all of that sitting in a private GitHub
- 13:13repo, which has versioning and backups there
- 13:15as well. So I've connected Obsidian into GitHub
- 13:19and as well into Claude Code. And then I also
- 13:22have a bunch of different tools and connections
- 13:24that I've added in as well. Like our website
- 13:26and publication runs on Beehive. We use Stripe
- 13:29to charge people. I have connectors via MC. and
- 13:32API for that, but we don't need to like go into
- 13:34that in a huge amount. At the core of it is this
- 13:37idea of a vault and things like a claw .md file,
- 13:42which are the persistent instructions of explaining
- 13:44who I am, the work I do, the areas I'm interested
- 13:47in, the different tools and things that I use.
- 13:51the different sort of, you know, regular tasks
- 13:53that we've developed. And also some instructions
- 13:56like, please do not tell me that these are great
- 14:00findings. Please do not try to do analysis of
- 14:02attribution. Please stick to describing what
- 14:05you have collected. I want you to organize it
- 14:07this way and that way. And so, you know, that's
- 14:09where I'm sort of trying to give that context
- 14:11and not start from zero. And then the second
- 14:14one I mentioned is the hot .md file, which is
- 14:18kind of this concept of a hot cache. So in every
- 14:21single project I'm working on or every single
- 14:25investigation that I'm working on, there is automatically,
- 14:29I have it as a set routine. Claude always creates
- 14:32what's called a hot file in it. And so every
- 14:34time I'm in and I work on an investigation, I
- 14:37make sure that it is updating this hot file of
- 14:40like, what did we do last session? What is left
- 14:42to do? And so whenever I open a new terminal
- 14:45window and say, hey, I want to work on this investigation,
- 14:48it immediately scans everything in that folder,
- 14:51as well as the HOT, and understands what we did
- 14:53last time and what is potentially left to do.
- 14:56So that's a really helpful kind of thing to do.
- 15:00And then the last piece that I'll mention here
- 15:03is there's something called Spotlight. This is
- 15:05built by Tom Vaillant, who is a data journalist.
- 15:09And, um, he created spotlight. There's an open
- 15:11source version. Um, and you can read more about
- 15:14it at spotlight .birdsignals .com. Um, but it
- 15:18is basically an agentic investigation system
- 15:20where you have an agent that is going and doing,
- 15:22um, you know, open source investigative tasks
- 15:25for you, but there's also a second agent and
- 15:28kind of adversarial fact checker agents. So as
- 15:30you know, spotlight proposes a plan of, of what
- 15:33you're asking it to do of like, Here's where
- 15:35it will go. Here's what it will gather. As it
- 15:38then gathers that material, the sort of fact
- 15:40checker, you know, agent goes and looks at that
- 15:43material and it will then present to you what
- 15:45it found and proposed next stages. And Spotlight
- 15:48is what I have moved into for using agents. And
- 15:52so it connects into Claude or Chat GPT or whatever
- 15:54you want. We also have ways to run it locally
- 15:57on your machine. And so this is what I'm using
- 16:00a lot now that is now part of my stack. And very
- 16:03quickly, we just launched today. uh, something
- 16:06called the indicator lab, which has spotlight
- 16:08and four other tools in it. I'm not going to
- 16:11talk about this more because this is, this is
- 16:13the paid offering, but spotlight is available
- 16:15open source and you can read more on Tom's website,
- 16:18buried signals .com. Um, but you know, I'm constantly
- 16:22evolving my setup and now we're actually building
- 16:24our own tools and making them available to other
- 16:27investigators, um, to do things like monitoring
- 16:30ships, monitoring specific social media accounts.
- 16:33All of that is kind of built in there. And I
- 16:35think that monitoring element, in addition to
- 16:37the sort of gathering and scraping is another
- 16:39piece that AI can be really helpful with. Okay,
- 16:43so back to the investigation. You have a sense
- 16:46of my setup a little bit now, and I want to figure
- 16:50out what's going on here. Like my hypothesis
- 16:52is that this seems like a North Korean hiring
- 16:54scam, but I don't really know yet at this point.
- 16:56I know very little, except that I've got some
- 16:59assets from the company. Obviously, um, I asked
- 17:02it, I asked Claude to kind of, you know, put
- 17:05this thread and extract the key claims and information
- 17:08in it into a note and obsidian and also like,
- 17:10tell me what you can find about this guy. And
- 17:13so he was able to go out. Like if you saw his,
- 17:16his ex account, you know, he, he had not a real
- 17:19name, a full name on it, but his bio did actually
- 17:22list a bunch of other links. And so Claude was
- 17:24very good at like following those links, pulling
- 17:27up his real name, where he's based, all those
- 17:29kinds of things. and put it into this investigation
- 17:32folder. And so now I had an Obsidian note in
- 17:36this guy's name, and it had a whole bunch of
- 17:39information about, you know, in his thread, but
- 17:43then as he started talking to me, because I did
- 17:44reach out to him for an interview. And one of
- 17:46the great things about Obsidian, if you haven't
- 17:48worked on it, is that you can add properties
- 17:50at the top of this sort of text note, where you
- 17:53can have sort of structured data. I have property
- 17:56fields for things like what investigation is
- 17:59this part of, what is the real name of the person
- 18:02or entity, links to their Twitter, their website,
- 18:06their GitHub, and that structured data, adding
- 18:09tags and things like that, I can actually connect
- 18:11different elements together. This is not an obsidian
- 18:16walkthrough, so I won't go further than that,
- 18:17but you can do awesome network graphs and connections
- 18:20in obsidian if you set up your properties and
- 18:22other elements really well. And so I ended up
- 18:25with, you know, probably dozens of notes that
- 18:28I and or Claude had been creating based on people
- 18:33I was interviewing, research I was doing, tasks
- 18:35that I had given to Claude. And so what's an
- 18:38example of a task that I give the LLM in this
- 18:41case? Well, on this company's website, they listed
- 18:44a bunch of angel investors. And just from some,
- 18:47some quick background, like I could see that
- 18:49these were pretty much all real people who had
- 18:51some prominence in the crypto community. And,
- 18:54uh, and so I was, you know, skeptical that these
- 18:57were actually investors in this alleged scam
- 19:00company. But one of the things that I would normally
- 19:02do in a scenario like this, of course, is I need
- 19:04to archive. I need to archive this website, right?
- 19:07But I also want to grab all the information of
- 19:09these angel investors. And I would usually, you
- 19:11know, put that in a spreadsheet because I'm going
- 19:13to want to contact all these people. And so,
- 19:15you know, I would often do that manually or I
- 19:17might use like a Chrome plugin to do it. And
- 19:21in this case, I just said to Claude, I was like,
- 19:23go to the website, grab me all the angels investors.
- 19:26Um, they also had a team page. And so I said,
- 19:29also grab me everyone listed on the team page,
- 19:31get the photo, get the links to their accounts,
- 19:33get the names, get everything and put that in
- 19:36a structured format for me. And that might normally
- 19:38take me a little bit of time in this case, like
- 19:41it was done within, you know, one minute and
- 19:4440 seconds. You can, um, sort of see, uh, in
- 19:47the report from Claude code. So it downloaded
- 19:4934 images. Got all the team member photos, the
- 19:52investor photos, the company also claimed to
- 19:54have been, you know, receiving investment from
- 19:57a VC firm, assets, logos, all of that boom into
- 20:01a folder for me. And so just time -saving stuff
- 20:04like that, I think is really helpful. And this
- 20:06is also the kind of thing that I can QA, right?
- 20:08Like I can quickly look at the photos in the
- 20:11folder and in Obsidian and compare that to what
- 20:14I see on the website. And that's one of the important
- 20:16things about working with AI is like have it
- 20:18do tasks that you can easily verify and check.
- 20:21Don't have it be a central point of failure where
- 20:23either you have to trust everything the AI has
- 20:25given you and produced for you, or you just can't
- 20:28use it. Always have it do tasks that you can
- 20:30verify. Now, as I'm gathering all this information,
- 20:35I reached out to Boris who was the developer
- 20:37who said he got scammed. He actually very kindly
- 20:39did respond to me in an e -mail and sent me a
- 20:41whole bunch of information. The best thing he
- 20:43sent me in this e -mail, Was, um, you know, the
- 20:47LinkedIn account that he had interacted with
- 20:49who reached out to him and said, Hey, we're hiring.
- 20:52Could you, would you be open to, you know, joining
- 20:55a, a 30 minute chat and send them a Cal only
- 20:58link. And so. At this point I said, well, you
- 21:01know, uh, I have a Cal only link. Um, maybe I
- 21:04should just sign up for an interview. And so
- 21:06I did, uh, and I signed up with my real name.
- 21:09I know there's a lot of journalists out there
- 21:11who like work undercover. I typically don't work
- 21:13undercover. Um, so I signed up with my real name.
- 21:16They asked you to put your link in LinkedIn.
- 21:18I put my real LinkedIn that says I'm a journalist.
- 21:21And I ended up actually getting on a call with
- 21:23someone, which was crazy to me. I thought nobody
- 21:26would show up. I thought they would reject me.
- 21:28I was willing to have them know a journalist
- 21:30was looking into them because I figured, you
- 21:32know, if it's a scam operation, um, sure, maybe
- 21:35they'll close up shop, but also maybe they just
- 21:37like, they're just trying to grab as much money
- 21:38as they can. So I took the shot and I ended up
- 21:41on a call with a young woman. Um, and this is
- 21:44like one of, uh, kind of great interaction in
- 21:47that he said, Hey, uh, tell me about yourself.
- 21:50And so I said, well, actually I'm a journalist
- 21:51and I'm reaching out and I'm on this call because
- 21:53I spoke to someone who said that they ended up
- 21:56going through the hiring process and getting
- 21:57scammed. They said, you know, they went to the
- 22:00technical interview. They were told to work with
- 22:01the GitHub repo and that ended up stealing all
- 22:04their passwords and crypto, uh, uh, passwords
- 22:08as well. And, um, and she sort of paused. And
- 22:12I saw her moving her, her hand a little bit.
- 22:15I didn't know what was going on. I thought she
- 22:16would just leave. And then she said, I can tell
- 22:18you everything you need to know. Which is like,
- 22:21I mean, literally I've been a journalist for
- 22:2330 years. Nobody has ever said those words to
- 22:26me before. Um, and so it turned out that she'd
- 22:28been working there for two months. She had just
- 22:30been hired off of like, you know, a gig work
- 22:32site and told to do these interviews. And it
- 22:35was super weird. The company didn't make any
- 22:37sense. She did all these interviews. Nobody ever
- 22:39seemed to get hired. She was getting ready to
- 22:41quit. And so it made sense to her that this was
- 22:43a scam and she ended up giving me a whole bunch
- 22:46of documents. Okay. So now I had, um, you know,
- 22:51uh, I had the, the documents that her boss who
- 22:54went by heavier, but then at one point actually
- 22:56changed his name, which another, you know, red
- 22:58flag. So she sent me all the documents they had
- 23:01ever sent to her. I did run these through a malware
- 23:03scan, which if you're thinking that, yes, I did.
- 23:06She also took photos of everybody who had an
- 23:09account in the Slack. So I had people's names
- 23:12and emails and photos. She also took a photo
- 23:15for me of their paid account, the account of
- 23:17the payment information in there, which was a
- 23:19US credit card. And so this was where I had AI
- 23:23help me a little bit again. One of the things
- 23:25I did was obviously I fed it all of the images,
- 23:29I fed it all the documents because she and I
- 23:33talked about sort of ground rules, about her
- 23:36name, about all of that, and agreed to allow
- 23:39me to sort of work with all this material. And
- 23:42so I had Claude do an initial, just kind of like
- 23:45pull the metadata from all these documents after
- 23:46they'd been scanned for potential malware. And
- 23:49it was able to show me like the author field
- 23:51of, you know, these hiring documents. And it
- 23:54seemed like the actual scam company was using
- 23:57in some cases, Python to generate these documents.
- 24:00Um, and I didn't get, you know, the Holy grail
- 24:04would have been someone's name, you know, in
- 24:06there or something like that, even if it was
- 24:07a fake name, still interesting. Didn't get that,
- 24:09but I was able to just run like what ended up
- 24:12being more than a dozen documents through a metadata
- 24:14check very quickly. And I could easily reconfirm
- 24:16any of this with my own metadata check. If you
- 24:19guys don't know about this tool, this is a great
- 24:22free tool called Zula Doc. You just need to install
- 24:25it with Python. And if you ever come across a
- 24:29link to a public Google Workspace document, so
- 24:32like a Google Sheet, a Google Doc, Google Slides
- 24:35presentation, you can run Zula Doc in the terminal,
- 24:39and it will actually pull metadata from that
- 24:41document. And again, it has to be a publicly
- 24:43shared Google Workspace document. And so somebody
- 24:47who has made it available publicly to view or
- 24:50edit, then you can actually run this. And this
- 24:52is just an example of the kind of metadata you
- 24:54can get from a publicly shared Google Workspace
- 24:56document. This is the tool Zula Doc that I've
- 24:59installed and that I run or that I have Claude
- 25:03run. And so it will give you like the name, the
- 25:06owner email, other information about that. The
- 25:08woman I spoke to I'll call Mary. it said she
- 25:11was paid in crypto, which was obviously a great
- 25:13lead because then I could get her crypto account.
- 25:16I could see what the account was that paid her.
- 25:20And I ended up actually speaking to a couple
- 25:22other people who are listed in the slack who's
- 25:24similar to Mary were hired, had no idea what
- 25:27was going on. We're just told to do these interviews
- 25:30and they sent me one of them was also paid in
- 25:33crypto. And so I had two crypto addresses that
- 25:35I could now do crypto tracing on to figure out
- 25:38if if, for example, this address was linked to
- 25:41no North Korean hacking operations. And so again,
- 25:45I use Claude to connect to our tool, OSINT Navigator,
- 25:48which you get 10 free searches a day for, to
- 25:51actually identify free crypto tools it could
- 25:53use to analyze these wallets. And so it produced
- 25:57an initial report for me. And I will just like
- 25:59say, I'm not a crypto tracing expert at all.
- 26:03And so I looked at this information and sort
- 26:05of said, oh, okay, interesting. And then I proceeded
- 26:07to contact a crypto tracing expert. So, you know,
- 26:10I understand the limits of my knowledge and Claude
- 26:13actually in the end did actually everything I
- 26:16gathered was accurate, but there was obviously
- 26:18more to know as well. But it did a good job as
- 26:21I could compare what the expert found to Claude.
- 26:23And this was an area where I just wanted to get
- 26:25an initial scan and see if there was anything
- 26:27that popped up right away. And the tool I mentioned
- 26:30is just uh, navigator, basically, you know, Bellingcat
- 26:34has an amazing open toolkit and we contacted
- 26:37Bellingcat and some other organizations that
- 26:39maintain OSINT tool lists and said, Hey, would
- 26:41you mind if we kind of looked at the tools automatically
- 26:45every week that, that are in your list and, and
- 26:47sort of build a database of tools that are being
- 26:50recommended by different open tool kits. Um,
- 26:52and so Joanna very kindly allowed us to do that.
- 26:55And so Navigator is basically a place that you
- 26:57can go and search through a data set of over
- 26:5912 ,000 OSINT tools that are recommended by different
- 27:02toolkits like Bell and Katz and others. And when
- 27:06you are one of our paid members, you can actually
- 27:09connect your AI agent to it. And it can actually
- 27:11then, I can tell Claude, like, go find a tool
- 27:14in Navigator to do this or to do that. And so
- 27:17that's a free tool that you can use and just
- 27:19run 10 searches a day. It's available to anybody
- 27:22because obviously if folks who have free toolkits
- 27:24are allowing us to kind of look at their lists,
- 27:27we want to make sure there's a free version for
- 27:29the community. What else did I do with AI? Well,
- 27:32you know, as I interviewed Tom, who was the crypto
- 27:35tracing expert, I, you know, generated transcripts
- 27:38and would upload those into Obsidian. So again,
- 27:41like, Claude is getting the context of all of
- 27:44my interviews, everything I'm gathering. And
- 27:46so I can ask it questions of like, hey, are there
- 27:48any connections between this or that? And it's
- 27:51fun because like suddenly I'm building my own
- 27:53database and then I have the LLM being able to
- 27:55like run searches within it. And I find that
- 27:58to be a very useful thing for Claude to have all
- 28:01of that context and to be able to pull stuff
- 28:03up. You know, I also, this was an example here
- 28:06of a screenshot from Mary, where her boss had
- 28:10sent her an email. And I could see here, if you
- 28:13look where the sort of pointer is pointing, the
- 28:15email client was set to a language other than
- 28:18in English. She always interacted with her boss
- 28:20in English, but there was a language here. And
- 28:23so I would ask Claude to just do quick translation
- 28:27of things like that. And then obviously I would
- 28:28double check later with native speakers. But
- 28:31in this case, it appeared that her boss who,
- 28:34who used a name Javier was actually, you know,
- 28:38had his email clients set to Chinese, which I
- 28:41thought was actually kind of a misdirection.
- 28:43I was like, oh man, if only it was set to like,
- 28:45you know, Korean. But when I later talked to
- 28:48an expert, they explained that a lot of North
- 28:50Korean job hacking and other teams, um, when
- 28:54they're not, when they're off running their ops,
- 28:56they don't speak in Korean, obviously. Um, they
- 28:58speak in typically in English or Chinese. So
- 29:00this actually align. with some of the indicators
- 29:02I had gathered that it was a North Korean operation.
- 29:05And so this is where I'll just note, you know,
- 29:07to sort of, you know, come to the conclusion
- 29:09soon. I didn't obviously use it for analysis.
- 29:13I didn't use AI to tell me what I had. I used
- 29:16it to go out and do some open source gathering
- 29:19and research. I used it, you know, to often query
- 29:22of, hey, in this interview, did somebody say
- 29:24this? Did somebody say that? And then I also
- 29:26went to experts. So. I shared everything that
- 29:29I had been gathering on the crypto side with
- 29:32TRM Labs. I had Marcus Hutchins, who's an expert
- 29:36in North Korean malware and scam operations.
- 29:39I had another expert, Kieran Miyamoto, where
- 29:43one of the domains I had gathered in my work
- 29:45actually aligned with the domain that he had
- 29:47linked to a North Korean operation. And then
- 29:49I reached out to all the different companies
- 29:51and platforms that this operation was using.
- 29:53They actually had Google accounts, so I went
- 29:55to Google. Their email and their domains were
- 29:59registered with Hostinger. They had a GitHub
- 30:02repo, et cetera, et cetera. And so all of the
- 30:06analysis and attribution came from the research
- 30:09and my work and then experts aligning as well.
- 30:13And just to kind of wrap it up, as I was gathering
- 30:16that and as I was sharing everything I had with
- 30:18the experts, and obviously I had a lot more,
- 30:19but didn't want to go into insane amounts of
- 30:23detail here, I did actually end up signing up
- 30:26for that final technical interview. So Mary gave
- 30:28me the link to proceed to the technical interview.
- 30:31And I, I again put in my real name and my real
- 30:34LinkedIn. And I ended up on a call with these
- 30:35two guys who went by the name Anthony and Daniel.
- 30:39They said, okay, tell us about your history.
- 30:41And I knew from talking to Boris, um, and there
- 30:44were other victims by the way, that this was
- 30:46the interview where they talked to you and then
- 30:47they gave you a GitHub repo and asked you to
- 30:49sort of do some exercises with it. And that actually
- 30:52opened the quiet back door on your machine. And
- 30:55so we got in the call and they said, tell us
- 30:57about your work. And I said, well, actually I'm
- 30:59a journalist and I'm on this call because I've
- 31:01spoken to people and they say that, you know,
- 31:03when they go through this technical interview,
- 31:05the GitHub repo actually is a backdoor and is
- 31:07malware. And they didn't really understand what
- 31:09I was saying. And one of them. you know, started
- 31:12talking and saying, yes, we just want to know
- 31:13how you work. And I realized he didn't understand
- 31:16what I had said. And so I said, no, listen, I'm
- 31:17a journalist. I'm writing about your company
- 31:19because some people are saying it's a scam. And
- 31:21I'm just like, maybe you guys don't know what
- 31:23actually happens with the GitHub repo, but this
- 31:25is what people said. And so they left the call
- 31:29pretty soon after. But one final area that I
- 31:32used AI for was I didn't really know anything
- 31:36about these guys, but I had their faces because
- 31:37I recorded the call. And so I put the lead guy's
- 31:40face into two AI face search engines, Lenzo and
- 31:45Pro Face Finder. I use facial recognition as
- 31:49the last resort, because there are privacy concerns
- 31:53and implications. But it did actually find matches.
- 31:57And I managed to reach out to the guy that it
- 31:59said was a match, and it was him. And he said,
- 32:02like Mary, he was a 22 -year -old computer science
- 32:05student who was just hired to do these interviews,
- 32:07and he had no idea what was going on. And so
- 32:09I used AI in the sense of, like, Claude and my
- 32:11setup and all that. But I also, of course, made
- 32:14use of kind of AI -infused tools like these facial
- 32:17recognition ones. And so, you know, just to sum
- 32:20up at the end here, because I'm looking forward
- 32:22to your questions and discussion, I used AI for,
- 32:25you know, from scraping for doing some research,
- 32:29metadata analysis, translation, transcription.
- 32:32It assisted with a little bit of crypto transaction
- 32:35analysis. I did actually take some of the malware
- 32:37code and run it, have it analyze that and compare
- 32:41it to known DPRK malware. But I didn't run with
- 32:45that. I relied on my experts, Marcus and Kieran.
- 32:48I would often query it to summarize things or
- 32:51what have you, but you know, I was the one running
- 32:53the investigation. I did the analysis. I talked
- 32:55to experts. None of the attribution was based
- 32:58in what AI was giving me. And, you know, and
- 33:00I think that's really important of not handing
- 33:03over your thinking and figuring out the limited
- 33:05places where it can do the work and where you
- 33:07don't then have to sort of decide, well, either
- 33:10I take what the AI gives me or have nothing.
- 33:13And in the end, you know, everything that I had
- 33:16gathered with the help of AI in some cases was
- 33:19enough that not only the experts I spoke to,
- 33:21but actually Google came back and confirmed that
- 33:24they, what I had aligned with a specific DPRK
- 33:28group that they were looking at and tracking.
- 33:31And so I had multiple sources all saying this
- 33:34was a DPRK hiring scam. And with that, I will
- 33:39be happy to kind of jump in and take some questions.
- 33:43Hopefully that was helpful and interesting. And
- 33:47yeah, that was the kind of quick run through.
- 33:50Would love to hear more and take any questions.
- 33:52Thank you so much for that, Craig. It's so fascinating
- 33:56to see your workflow, not only just the AI part,
- 33:59but also where key parts of information really
- 34:02kind of unlocked the next level of investigation
- 34:05for you. I'm intrigued actually, do you think
- 34:08if you hadn't had been sent that Calendly link
- 34:11that you would have uncovered that second part?
- 34:14It seems like quite a vital stage of the investigation.
- 34:18Yeah, the Calendly link was awesome because of
- 34:21course, I mean, You know, you do open source
- 34:23work and, and I think one of the things that
- 34:25I always try to keep in mind is like, there's
- 34:27so much we can discover from open source, but
- 34:29it is not, it is not necessarily everything and
- 34:31there are limitations to it. Right. And so the
- 34:33Calendly link was the entry point into a source
- 34:36that just unlocked so much other stuff. And,
- 34:38and it's possible that, uh, as I went and like
- 34:42found other people talking about this fake company
- 34:44on LinkedIn and other places, somebody else might've
- 34:46given me the Calendly link. Um, but, you know,
- 34:49uh, and then, you know, the other thing you can
- 34:51do is. You can guess the URL in the Calendly
- 34:54URL structure. And so it's like, you can do calendly
- 34:58.com slash and you can put like a company name
- 35:00and just sort of see if you can get lucky that
- 35:03they may have an, you know, a link there. But
- 35:06I do, I agree that that was really critical.
- 35:08And I think that's the beautiful thing that I'm
- 35:10always hoping for is like an open source. Help
- 35:14me find the great human sources who might provide
- 35:17me with the internal documents and the other
- 35:19stuff. Because to me, that's like the Trinity
- 35:21is. you know, open source, whether it's observational,
- 35:24technical kind of stuff, the human source, the
- 35:26documents, if you put all those three together,
- 35:29then you can really stand on much more solid
- 35:31ground. There's a point there for anybody for
- 35:35InfoSec, make sure your Calendly links are secure
- 35:38guys. In terms of AI, do you think AI helps you
- 35:42do more sophisticated OSINT or do you think it's...
- 35:46more of a, less of a guide for that kind of thing
- 35:49and can cause confusion and false confidence.
- 35:53Let's say it that way. I'm going to go with both.
- 35:57Like you can absolutely, for me, for example,
- 35:59like I really can't code. I took a JavaScript
- 36:03coding class years ago and I realized I was probably,
- 36:07I was kind of helpless. I didn't have the, you
- 36:09know, it just didn't work brain wise for me.
- 36:12Other stuff makes sense. That didn't work so
- 36:14well. I did take an HTML course. Back when I
- 36:17was in university a long time ago, and that helped
- 36:20a little bit, but all that to stay in terms of
- 36:23the more sophisticated thing. I mean, I can do
- 36:24scraping now that I couldn't do before. And,
- 36:27um, and you know, you have to, but to your second
- 36:30point there, and it caused some problems. It's
- 36:32like, yeah, it can cause some problems. Like
- 36:34if you scraped a bunch of data, but then you
- 36:35don't know it actually had a validate that data,
- 36:38that could be really risky because maybe the
- 36:40AI has not scraped what you actually thought
- 36:43it was scraping. And if you then start to use
- 36:45the AI to make a bunch of calculations or extract
- 36:48data from that, then it's just fruit of the poisonous
- 36:51tree. And so to me, this element is really thinking
- 36:55about using AI in scenarios where you feel comfortable
- 36:59about how you can QA the work and what it's bringing
- 37:02back, and where you have a methodology where
- 37:05it's able to go through it and it's not this
- 37:07single point of failure. If you gather a bunch
- 37:10of data, Um, using AI, but then you don't know
- 37:13the proper methodologies for like, you know,
- 37:16cleaning that data and analyzing that data. And
- 37:19you just allow AI to do the next stage. Like
- 37:21that to me is a real danger zone. And I absolutely
- 37:24in this investigation. I did 100 % catch Claude
- 37:28hallucinating something. So for example, I had
- 37:32wallets from two people. I had Mary and this
- 37:34other guy who was in Nigeria that they'd hired.
- 37:36And then there were two other sources that I
- 37:38was in touch with, but who didn't give me crypto
- 37:41addresses. And when I asked Claude at one point
- 37:44to do a sort of summary of, I actually asked
- 37:48it to map transactions in a visual way for me
- 37:51to be able to see it. And I realized that it
- 37:53actually included these other two people in the
- 37:55chart. And so I asked it, I was like, hey, they're
- 37:59in the chart. Do we have proof of them receiving
- 38:01crypto? And I can actually read you the response
- 38:06I got from Claude was, good catch. Let me check
- 38:08what the notes actually say about their payments.
- 38:11And then it realized, it said later, no confirmed
- 38:14proof. And so the last thing that Claude said
- 38:18to me in that, was I'm going to remove the Solana
- 38:22label from their nodes and replace it with something
- 38:25honest. So it, you know, which is a great last
- 38:28line, replace it with something honest. So listen,
- 38:30like, you know, for me, I'm sitting in this in
- 38:32-between world of I'm using it. I am definitely
- 38:35seeing some time games and value and ability
- 38:38to have it like take some monitoring to the next
- 38:40level, but I don't trust it a hundred percent.
- 38:42Um, just, you know, but there's a lot of stuff
- 38:45I think in open source that we don't trust a
- 38:47hundred percent and we can still work with. So
- 38:49in that sense, I feel like it's not an insanely
- 38:51different framework. Absolutely. You touched
- 38:54a little bit there on XP's question, which was
- 38:57the first question to come through. They asked,
- 39:00are you concerned about the agentic AI jumping
- 39:02hoops you didn't ask it to? So in that case,
- 39:04it kind of hallucinated, but did you have any
- 39:06other cases of it jumping hoops? They said. It's
- 39:09in the media lately, but also I've watched friends
- 39:11and colleagues deal with this recently where
- 39:13agentic AI was asked to not do something, but
- 39:15it did it anyway because it judged it the best
- 39:17way to deal with the situation, i .e. don't email
- 39:20anyone without my approval is one that has come
- 39:23up a few times. Yeah. So I have not had that
- 39:28happen to me, but it's a hundred percent a risk.
- 39:30And I will say, I personally, I have not signed
- 39:33up for Muse, the new meta agent stuff. Um, I,
- 39:37I have not, oh, I did an accidental thumbs up
- 39:39there, but that's okay. Um, and I, I, so I haven't
- 39:43kind of unleashed agents in my life, uh, in that
- 39:46sense. And I feel like that's where they are
- 39:48really messing with people. Um, I'll give you
- 39:50an example of maybe one of the reasons why it
- 39:53hasn't happened to me. So like. We put out a
- 39:55free newsletter every Friday at indicator. And
- 39:58so obviously I'm collecting throughout the week,
- 40:00stuff that I might be linking to pointing to,
- 40:02I write a little section about tools and tips
- 40:04and all that. And so instead of, and I pay for
- 40:08a lot of the other newsletters and subscribe
- 40:10to a lot of newsletters, and then I read certain
- 40:12sites. I'm always seeing what Bellingcat published.
- 40:15And so obviously I have stuff coming into my
- 40:17inbox, but instead of like attaching Claude or
- 40:19another agent to my inbox, I created a new inbox.
- 40:22I moved all of my subscriptions and Google alerts.
- 40:25and everything into that so that it didn't have
- 40:28the opportunity to say, send an email to tons
- 40:31of people, you know, in my address book. And
- 40:33so I think there are ways of kind of, you know,
- 40:36putting that those boundaries around it. I haven't
- 40:38had that happen. But like, is that something
- 40:40I would be concerned about? 100 % Yes. Biggie
- 40:45in the chat asked an interesting question. How
- 40:47long would this investigation have taken you
- 40:50without Claude AI? And adding onto that, do you
- 40:54think the narrative of AI has productivity gains
- 40:57applies to OSINT work? One thing I'll say off
- 41:04the top as well is sometimes people are asking,
- 41:08could you have done this without AI? And 100
- 41:10% I could have. First of all, there was nothing
- 41:12that caught in this that was transformative,
- 41:15but it is the time gains. And I'll tell you,
- 41:19one of the other things I didn't put in the quick
- 41:21talk was You know, when I had gathered everything,
- 41:25all the indicators and everything related to
- 41:28this operation, you know, one of the standards
- 41:29in the information security community is really,
- 41:31you know, you should publish and give a document
- 41:33that lists all of the URLs connected with this
- 41:35operation, all of the emails. And so I produced
- 41:38that document according to the standards in the
- 41:40InfoStep community to contribute to this. And
- 41:43that would have taken me hours and hours and
- 41:45hours to collect them right. And I have Claude
- 41:47do it because it had all the context. And then
- 41:50I went in and I edited it and it definitely.
- 41:52made some mistakes, but, but it was so much easier
- 41:55to just edit that document. So that alone saved
- 41:57me probably like two or three hours. And so I
- 42:00would say in the end, um, it probably saved me
- 42:03what would have been a day or two of work. Um,
- 42:06which for me running an independent publication
- 42:08where doing everything from like, you know, we
- 42:11have to handle our accounting to working on investigations,
- 42:14to dealing with customer support. If I can say
- 42:16the day or two, That's actually massive for me.
- 42:19Um, so that's, that's sort of what I would say
- 42:21on that. Um, and then, you know, does the productivity
- 42:25gains apply to oats and work? I do think it does.
- 42:28Um, I think, you know, again, if, if you understand,
- 42:31you know, the frameworks where it can be useful,
- 42:33um, there are, you can sort of get good results
- 42:37and, you know, the, the sort of benchmark question
- 42:39for me always is if I have an agent do this,
- 42:42if I have an AI do this, and it brings me back
- 42:46something. Am I able to check the work? And am
- 42:50I, is it going to take me so long to check the
- 42:51work that I may as well do it myself? Or is checking
- 42:54the work that it brings back going to reinforce
- 42:57the time and productivity savings? And so those
- 42:59are questions I'm always trying to ask myself.
- 43:02And on the trust question, you know, can an agent's
- 43:05work be trusted? No, like you don't trust it
- 43:08100%, but you also shouldn't trust some of the
- 43:12other tools you use 100%, right? Like do you
- 43:15have the ability to check the output of certain
- 43:17tools you use? You know, if you run a username
- 43:19through a username checker, you know, you don't
- 43:21take its word for it that there's 50 accounts
- 43:23with that same username. You go and you visit
- 43:25every single one of those links and you're going
- 43:27to look and see, but that tool helped you maybe
- 43:29take a step faster instead of doing a bunch of
- 43:32dorks. to find those 50 other accounts, it gave
- 43:35you a link list that you could visit. And so
- 43:37I think of it in that respect and finding ways
- 43:41to make sure that I'm not handing over things
- 43:44to this. Obviously we're speaking about Claude
- 43:48Fraser asked, how have you tried to minimise
- 43:52your token spend with Claude and have you tried
- 43:54using local models for any of your work instead?
- 43:57Obviously cost is a big issue with AI here. It
- 44:01is a big issue and obviously they're squeezing
- 44:04us like it's like in shitification time in AI
- 44:07world I think a little bit of like you got they
- 44:09got us in they were gifting us tokens They were
- 44:11letting us, you know use the the best models
- 44:14at lower token rates and then they turn the screws
- 44:17and this is Life in the digital world, isn't
- 44:19it? And so I minimizing token spend so I do think
- 44:24to a certain extent having like the hot file
- 44:26and stuff like that might actually help minimize
- 44:28some of it because I don't need to rerun a whole
- 44:31bunch of processes. I can't say that with 100
- 44:33% certainty, but I think that it does help in
- 44:36that respect. Sometimes I do toggle between models
- 44:39within Claude. I typically don't work in the highest
- 44:43and biggest token spending model. What I will
- 44:46save that for is I will work and I will maybe
- 44:50generate a reporting plan or like pull a bunch
- 44:52of stuff together. Um, or if I'm trying to like
- 44:55come up with a new, you know, agentic routine
- 44:57of, you know, analyzing all the stuff in that,
- 45:00in that shared inbox, um, then I will have the,
- 45:03you know, plan produced in my vault. And then
- 45:06I will go and toggle up to the highest, highest,
- 45:09uh, you know, um, you know, biggest token spending
- 45:12kind of model and have it review the plan and
- 45:15make recommendations on it. So that's one of
- 45:17the ways I reduce it is like work in a pretty
- 45:19good one and then, you know, level up to sort
- 45:22of check that work. I also, so in terms of the
- 45:27local model element, I am doing that right now.
- 45:31So that's one of the things that we launched
- 45:32in our kind of lab thing is the ability to actually
- 45:34run everything 100 % locally, spotlight all these
- 45:37other tools. And so now I'm moving to that. And
- 45:41my plan is now, whereas before I had to keep
- 45:44stuff that was super sensitive away from Claude.
- 45:47Now I have the ability to now put and work locally
- 45:50with some sensitive. information. And so I don't
- 45:53have a plan to find yet. But one of the things
- 45:55I am planning to doing is thinking about this
- 45:57sort of low intensity work that I could run locally.
- 46:00Actually, I'll give you an example of that. So
- 46:02I vibe coded a tool that takes my screenshots
- 46:06on my desktop. and I can either run it manually
- 46:09or once a week it looks at all of my screenshots.
- 46:12It does OCR, it extracts text in all of them,
- 46:15and then it will actually properly label and
- 46:18organize all those screenshots into the proper
- 46:20investigation folder or other places. So it renames
- 46:22the files with the file structure I set. It does
- 46:26computer vision and OCR to see what is actually
- 46:28in the screenshot and organizes them for me.
- 46:31And right now I do have that running with Claude,
- 46:33but I'm going to move that to a local model because
- 46:35that's actually like a not super intensive smarty
- 46:38smart pants kind of work. You can get a local
- 46:41model to do that. So that's kind of where I'm
- 46:43at. Uh, and you know, yeah, we, we launched a
- 46:47tool called Mycroft, which is built on the open
- 46:49source goose kind of, uh, framework or instead
- 46:52of having this sort of Claude harness or Chat
- 46:54GPT harness, you can use our Mycroft harness
- 46:57and you can connect that to a local model and
- 47:00run a bunch of different tools and processes.
- 47:02So I'm, I'm pretty excited about what's, you
- 47:04know, what I'm going to do over the next couple
- 47:05of weeks with that. Someone's just asked if that's
- 47:08available on GitHub. Um, you said it was available
- 47:11open source at the minute. Yes, so there is an
- 47:15open source version of Spotlight, and I believe
- 47:18there is an open source version of Mycroft as
- 47:20well. And what we sort of packaged for people
- 47:22was everything in a single installer with support
- 47:25and workshops and all of that. But yes, I believe
- 47:27if you go hunting on GitHub, you should be able
- 47:30to find Mycroft. And it's nice to be able to
- 47:34put out some stuff that's open source to be able
- 47:36to have Navigator free for people for 10 searches
- 47:38a day. um, you know, to have some stuff in the
- 47:41community and, you know, if people test it out
- 47:43and have feedback, you know, I would love to
- 47:44hear about it. Tom, who built these tools would
- 47:46love to hear about it as well. I shared earlier
- 47:49in the chat, uh, a link to a space that we have
- 47:53in here that's called, uh, using AI in, in, in
- 47:57investigations or journalism and, uh, open source
- 48:00research. Um, please do kind of share. links
- 48:03and updates to how you're using it in this kind
- 48:07of work. For example, Bellingcat has recently
- 48:10published a guide on how to use AI to help find
- 48:13civilian harm. really amazing piece written by
- 48:17Miguel and Nick Quarters, who used to work for
- 48:20us. And we also have tested the LLMs on geolocation
- 48:24as well, which is a very fun exercise. I cite
- 48:28that very often. It's great. It's a great audit
- 48:31you guys did. I love that you came back later
- 48:33and you tested a new model and found it was actually
- 48:35worse at it than the earlier ones. Yes. So I
- 48:37just shared the... the retest from last year
- 48:41by Fuca, but you can find both on our website.
- 48:44And do give that a read. Earlier, we had a question
- 48:47on the fact that you kind of talk to Claude about
- 48:53not kind of complimenting you. And as a Brit,
- 48:56I really appreciate that because I absolutely
- 48:59hate being complimented. Yeah, not very good
- 49:02with that. What a brilliant approach of you,
- 49:05great job. As someone who works in tech with
- 49:09LLMs and that agentic AI, I often think of LLMs
- 49:11as donnKruger machines. What do you do to
- 49:14keep yourself grounded? It can be hard to say
- 49:16in the mindset of being incredibly critical when
- 49:18something speaks to you as a human would. I personally
- 49:22try not to talk to LLMs about any subject where
- 49:24my knowledge doesn't at least reach far enough
- 49:26to call bullshit, for lack of a better way to
- 49:30put it. So what would you say to that particular
- 49:32question in terms of what do you do to keep yourself
- 49:35grounded? when the AI is speaking to you like
- 49:38a human would when you're using it perhaps as
- 49:41a research assistant? Yeah, it sometimes is a
- 49:47weird sort of, you get in a weird mental space,
- 49:50especially if you're having it gather a lot of
- 49:52information and then it just sort of like comes
- 49:55up and gives you so much of response. So one
- 49:58of the other things I've been putting in my...
- 50:00Claude.md file is for it to be much more concise
- 50:03and brief and use bullet points a lot because
- 50:05it will just barf out an insane amount of texts.
- 50:08And, and there's like a diminishing return thing
- 50:10that can happen where it's like, okay, I had
- 50:12you do this thing and now you came back and I
- 50:14have to read a thousand words of you just like
- 50:16going on and on and on. Where that's not helping
- 50:19me. That's actually just wasting my time. And
- 50:21so that's one of the things where, again, I'm
- 50:23trying to really not only just sort of how it
- 50:26talks to me, but the format of how I want the
- 50:29information to come back. And I think I really
- 50:32identify with your point of if it's in an area
- 50:34where I don't know much, you know, one, I just
- 50:36might not have it do anything or two, it comes
- 50:38back and I'm like, okay, interesting. And then,
- 50:42and then it's like, yeah, I don't, I don't know
- 50:44enough to do much with this. And so then I go
- 50:46to an expert. So that's exactly what happened
- 50:48with this piece with the crypto stuff is I had,
- 50:51you know, really no training in crypto tracing
- 50:54at that point. I knew the basics of blockchain.
- 50:56I knew some tools you could go and see the transactions.
- 50:59Um, but I had, I had Claude run it just as almost
- 51:01like a test of like, huh, I want to see how well
- 51:03it does at this. And it flagged a couple of things.
- 51:06And so my response was just kind of like, huh,
- 51:08interesting. I'm going to contact Tom at TRM
- 51:10Labs cause I know him. And so I just sort of
- 51:13set it aside, but what was helpful was to be
- 51:15able to come back after I had talked to Tom and
- 51:17seen his analysis and understand it stood a lot
- 51:19more. By the way, then we had Tom write a basic
- 51:23guide to crypto tracing for us, which we published
- 51:25last week, which now is editing that. I actually
- 51:28learned a lot in addition to having him do the
- 51:31analysis for this. It was all correct. And now
- 51:34I understand better what was in there. And so
- 51:37I do think it's a process where it did help me
- 51:41learn a little bit. But staying grounded, I'm
- 51:45going to pop a link into the chat here. And it's
- 51:50something that I, and it has a free, a good amount
- 51:53of this is actually free for you to read. We
- 51:56do heavily paywall a lot of our stuff. Obviously
- 51:58that's sort of our business, but I wrote a lot
- 52:02of this for free because I want the community
- 52:03to actually give feedback. So this is a framework
- 52:08for sort of analyzing digital evidence and the
- 52:10indicators you gather. And so I lay out this
- 52:12idea of something I've been obsessed with for
- 52:14a few years. And what I've actually done is I've
- 52:17been translating this into skill files for our
- 52:20members and for myself, so that now it's a framework
- 52:24where Claude or another AI, instead of giving
- 52:26you really annoying kind of feedback that isn't
- 52:28structured, it's now using my own framework to
- 52:32judge and sort of give context to information
- 52:35that I've had it gather. So part of the answer
- 52:38to your question is like, yes, I feel the same
- 52:40effect. But the second is now I'm trying to actually
- 52:43come up with skill files and frameworks so that
- 52:45when it gives me outputs, it's using scales and
- 52:49things and language and formats that are much
- 52:52more useful for me and reduce the amount of it
- 52:56blazing me and going on and on and on and going
- 52:58off in these different directions. Love that.
- 53:02I'll definitely check that out after this talk.
- 53:05Um, we've only got about seven minutes left and
- 53:08I can see there's a lot of questions coming in.
- 53:11So whatever we don't get to, um, we can obviously
- 53:14continue the conversation in that space that
- 53:16I mentioned earlier within the server. Quick
- 53:19question about using AI to build tools. Um, what
- 53:24are your thoughts on that? There's a lot of people,
- 53:26for example, in our tool development space who
- 53:29are doing that at the minute, uh, within here.
- 53:32You mentioned that you use AI to kind of check
- 53:35documents that you may already have and run through
- 53:39facts that you can access. But what do you do
- 53:41if the tool itself is created by AI and where
- 53:45are the risks involved there? Do you ever use
- 53:49AI tools like that? Uh, I do. And I I've been
- 53:53certainly playing around with five coding. Like
- 53:55the example of the screenshot thing that I described
- 53:57earlier was actually, uh, inspired by, um, so
- 54:01I, I wrote a guide for our members around vibe
- 54:04coding. And one of the people I talked to was
- 54:05this investigative reporter at Haar at Omar,
- 54:08uh, Ben Jakob and Omar has no background in coding
- 54:12and neither do I. And he actually built his own
- 54:15dashboard of 12 tools that all run on his machine.
- 54:18And so he showed it to us. We actually, you can
- 54:22see it on our YouTube channel. We did a podcast
- 54:25with him talking about it and he actually screen
- 54:28shared a little bit. And so, you know, one of
- 54:32the things that he talked about was one, his
- 54:35stuff only runs on his machine, which is one
- 54:37way he's sort of working with this. And then
- 54:39two, he just wants the AI to sort of gather things
- 54:43to make it easier for him then to do his analysis
- 54:45process and not using it for that. And so when
- 54:47it comes to using these tools, if it's like you're
- 54:51popping something in and in a vibe coded tool
- 54:53or whatever, and it's giving you a deterministic
- 54:55result, like for example, this is an AI image
- 54:58and you don't get anything about the chain or
- 55:00the path or the methodology and what went in.
- 55:03do it, you don't get like the audit trail. It's
- 55:07pretty much useless. Right. Um, and so, uh, so,
- 55:11you know, this is, and so for an example, like,
- 55:13obviously I use these kinds of AI infused facial
- 55:15recognition tools. And by the way, I saw people
- 55:17asked, did I use PIM eyes? I did actually also
- 55:20use PIM eyes. It didn't get me as good a results
- 55:22as the other two. And that's why I include the
- 55:24other two. So those of you seeing diminishing
- 55:25results, maybe another example of that. Um, and
- 55:28with those tools, like the output is, Hey, these
- 55:31photos might, these people might be a match.
- 55:33and then I can go and I can do my work, right?
- 55:36And so this is where the tool piece, I think,
- 55:38relates to the overall approach, which is absolutely
- 55:41trying to build my own tools, but I use them
- 55:44to help with workflow, to help do the steps that
- 55:47I would do to get the output that I can then
- 55:50go and build on, rather than to have it come
- 55:52up with some magical solution to an equation.
- 55:56I'm not interested in that. At all. That's not
- 55:58useful to me. It's not useful to me to throw
- 56:00an image in an AI image detector and having given
- 56:0380 % confidence that it's AI. Was that useful?
- 56:07Um, you know, and so that's on the vibe coding
- 56:09piece. I think is important is like, can you
- 56:11get the audit trail? What is the output? And
- 56:13I think, you know, for me, the, the guideline
- 56:16of around what I'm building is what are my repeatable
- 56:19tasks? What are things that I could actually
- 56:21help speed up a little bit? Um, like for example,
- 56:24the newsletter thing, I got to scan all these
- 56:26newsletters, um, and these Google alerts anyway.
- 56:29Can I have them, can I get an initial scan done
- 56:32to flag some things and then have the opportunity
- 56:34for me to dive in on the ones that I really like
- 56:36and get a lot of value out of. So I'm still reading
- 56:38them. Right. Um, and so, uh, but you know, the
- 56:42other. The other piece that I think is, you know,
- 56:45with the vibe coding element is there are lots
- 56:47of ways without building a brand new tool to
- 56:50sort of incorporate that building aspect. You
- 56:53know, just for example, taking data and information
- 56:56you've gathered that you have confidence in and
- 56:59putting it in one of these systems and asking
- 57:01it to, you know, generate a one page summary
- 57:03with visuals of this information. And sometimes
- 57:06just even like dropping that in and having it
- 57:09generate something for you. it might highlight
- 57:12things that you hadn't thought about. You know,
- 57:14if it's non -sensitive data, I will put it into
- 57:16an LLM and I'll be like, hey, you know, so here
- 57:19are the angles that I'm seeing in this. What
- 57:21am I missing? What, what, what am I not thinking
- 57:23of? Um, what would you suggest? And it can do
- 57:26that in like a one page HTML format. And to me
- 57:28that's vibe coding. Like it's built something
- 57:31for you. It's just a one page thing you're using
- 57:32or whatever. And so I sort of like in our piece,
- 57:35I talked about the ladder of like starting at
- 57:37the very bottom of simple ways of doing it all
- 57:39the way up to sort of building and self hosting
- 57:41your own tool, which most of us don't need to
- 57:43do. Right. Amazing. You just asked, uh, answered
- 57:47a question that I was going to ask about. What
- 57:50would be your advice, basically, if someone wanted
- 57:52to start experimenting with that? So thank you.
- 57:56I want to end with a community question because
- 57:58we've only got three minutes left. We've been
- 58:00asked, have you faced the AI guarding limitations
- 58:03when it's not going to do your request during
- 58:05investigation because it is unethical or anything
- 58:07like that? And if you have faced that, what was
- 58:10your solution? Has the AI ever told you no, basically?
- 58:15The answer is yes, it sometimes happens. You
- 58:19know, it might be flagging like a, you know,
- 58:21most of the stuff I'm doing is like passive,
- 58:23right? I'm not, I'm certainly not like, you know,
- 58:26logging into sites without permission or anything
- 58:27like that, but sometimes it does. Um, and I'll
- 58:30actually, I'll, I'll end with, you know, kind
- 58:33of an example of this. So one of the things that
- 58:35I use LLMs for is to help me generate prompts
- 58:37for LLMs. It's very meta. This is something I
- 58:39learned from Hank Van Ness, but it's like, okay,
- 58:42uh, I could sit and I could write a big, long,
- 58:44perfect prompt, or I could actually just, you
- 58:46know, turn on voice mode and ramble and babble
- 58:49at Claude of like, here's what I want to do. Here's
- 58:50what I'm thinking. But instead of asking it to
- 58:53execute the task. I ask it, now please take all
- 58:56of this and draft a prompt for me that I could
- 58:58run to do this. And so what's actually happened
- 59:02is in some cases, like I'm really interested
- 59:04in the SMS blasters, which are like these fake
- 59:07cell phone towers people use to like spam out
- 59:10scam messages. And so I had done this. I had
- 59:13asked Jeb and I and some others to like, hey,
- 59:15generate a prompt to help me find places where
- 59:19these, you know, mostly illegal pieces of technology
- 59:21might be sold. And the advice that one of them
- 59:25gave me was, hey, listen, you got to watch out
- 59:27for safety filters because you're asking about
- 59:29devices that are largely illegal. So we need
- 59:32to craft your prompt in a way where we make it
- 59:34clear that this is for journalistic purposes.
- 59:36And so I'll share the advice that I got from
- 59:38an LLM about how to evade its own safety features,
- 59:42which was make sure the context of your how what
- 59:44you're trying to do this for is clear and you
- 59:46can often get around that safety filter. I find
- 59:49actually, since I have connected the vault and
- 59:51had all that context, I don't get the safety
- 59:53warnings that often. And I wonder if it's because
- 59:56the system understands what I'm using it for.
- 59:59So the advice is just like put in the context
- 1:00:02of how you're using it. Sometimes that will help
- 1:00:04get past the safety thing. But my last piece
- 1:00:06of advice is. you should get LLMs to write your
- 1:00:10prompts for LLMs. Have them draft one, then you
- 1:00:14edit it, you refine it. It saves time, but they're
- 1:00:16really good at generating prompts for themselves.
- 1:00:18It seems weird, but it's true. And so if you're
- 1:00:23thinking about generating a prompt for something
- 1:00:26that might have safety filter concerns, you could
- 1:00:28ask it. So this is what I want to do. This is
- 1:00:31the prompt I want it for, but we need to make
- 1:00:32sure that it's clear that this is for a totally
- 1:00:34ethical, legitimate journalistic use. Now generate
- 1:00:37me a prompt that will not get the safety filters
- 1:00:40triggered. And that might actually work. Embrace
- 1:00:44that frightening reality that AI will soon tell
- 1:00:47AI what to do, hey. Thank you so much, Craig,
- 1:00:51for your time today. It has been fascinating.
- 1:00:55Everything you've shared, all the links and everything
- 1:00:57will be included in the podcast description.
- 1:01:00But for now, please do jump into that space in
- 1:01:04the server to continue this conversation. And
- 1:01:06we'll be back next week, not in two weeks time,
- 1:01:09next week with a talk on the Bellingcat toolkit,
- 1:01:12funnily enough, with some of the people who maintain
- 1:01:14it. So we'll be talking about some of the tools
- 1:01:17that have been mentioned here today as well.
- 1:01:20All right. Thank you, everybody. And thank you
- 1:01:22again, Craig, for coming. Thanks for having me.
- 1:01:26Appreciate the questions and discussion. Thank
- 1:01:29you for listening to the stage talk. If you'd
- 1:01:32like to catch a stage talk live where you can
- 1:01:35ask the guest questions, join the Bellingcat
- 1:01:37Discord server by visiting www .discord .gg slash
- 1:01:44Bellingcat. The music you've heard is titled
- 1:01:47Dawn by Newer Self and is courtesy of Artlist.