AI for automated OSINT reconnaissance - part one

Speaker 1: Welcome to the World of Intelligence, a podcast for you to discover the latest analysis of global military and security trends within the open source defense intelligence community. Now onto the episode with your host, Harry Kemsley.

Harry Kemsley: Hello, before we start this podcast episode, just a quick explanation that we're going to split it into two parts. So the first part we'll play now, and then we'll invite you back to join us for the second part very shortly. Hello and welcome to this edition of Janes World of Intelligence with me, Harry Kemsley, your host, and as usual my co- host, Sean Corbett. Hi, Sean.

Sean Corbett: Hi, Harry.

Harry Kemsley: Sean, we've been through two or three different podcasts around the topic of artificial intelligence. We spoke to Martin Keene and who was the other, Keith Dear. We spoke quite technically about it. I don't think what we've done to date was really address the key topic, which is, okay, great, so we have this technology, this advanced technology. How do we apply it to the open source environment? How do we apply it in a way that's actually effective? We also spoke to Harry, didn't we? Harry Lawson, who looked at large language models.

Sean Corbett: And the efficacy thereof. So it'd be really good to bring it all together from a practical perspective. And that's been the slight frustration that all these very clever people we've spoken to, we go, " Okay, how do we use it? How do we get most effect out of it?" And there's been lots of good chat and some real insight, but not really into the detail.

Harry Kemsley: So let's get to the detail of the practicalities. And for that, I'm delighted to invite to the conversation, Jim Clover. Jim, hello and welcome to the podcast.

Jim Clover: Hello. Hello, Sean. Hello, Harry.

Harry Kemsley: Jim, thanks for joining. Now, for the audience that doesn't know Jim, Jim's been around the intelligence world in a technical operational capacity for more than 20 years. He has a consultancy now, Varadius. Did I pronounce that correctly by the way, Jim?

Jim Clover: That'll be fine. Varadius, Varadius, I'm not too bothered.

Harry Kemsley: Thank you. I understand as a consultant, you help startups and enterprise- sized organizations with a similar topic of how you actually engage with the operational and technical environment, which I guess is based on your background of doing so in the intelligence community

Jim Clover: And certainly the last eight years in commercial as well as that 22 years in HMG. I think the combination of the two has been really helpful to have that exposure to non- government data challenges.

Harry Kemsley: Perfect. So given the topic that I've set up here for this conversation, Jim, you'll understand why you arriving and joining us for this is such an appropriate thing to do. It's a really good opportunity to discuss the actual application of AI. Before we get started though, could I just get, in your own words, what you think the letters AI actually mean? When we talk about the artificial intelligence capability, in your experience, what does it actually mean?

Jim Clover: Okay, so AI is, as I'm sure previous people that have attended this and talked about AI, it's been around for a long, long time. Neural networks were a bit of a black box to I guess most of the general public, not really understood, only by the very high- end mathematicians and people in AI, ML, machine learning and similar sort of practices. And I remember my first exposure to neural networks in the early'90s thinking, " What the hell is that?" And the guy tried to explain it to me and well, to be honest with you, I had no real clue what he was talking about, but it sounded really cool. So I wanted to keep an interest. So I've always had an interest in it. And to answer your question, for me, I think in the modern context of things, I think it was the dawn of the large language model for release to the general public, rather than being models that were used in very specialist fields and OpenAI must be given a lot of credit for bringing that to the general public. The ability for people to literally have a text box like Google search and I will touch on that as well. And the ability to ask a question from anything from the Roman Army through to the genesis of a certain molecule or the discovery of X, the ability to jump around subjects. So for me, really, AI in the modern context, and I like to view myself as almost like street AI in a lot of my commentary rather than deep scientific, because that's not my background. I'm a user of AI, and I do write software with AI and for AI, but I'm really interested in the impact it has for the person on the street and business in that sense. So for me, AI is very much about the democratization of knowledge, the democratization of code and coding, for the general public, really.

Harry Kemsley: I love that.

Jim Clover: And thus business, as well.

Harry Kemsley: I love the idea of a street AI practitioner because that really speaks very much to the practical application. The engagement is beyond the theoretical and technicals of which we've seen quite a lot in the conversations before, Sean, haven't we? Where we talked about the technology and what it does, how it does it, but the actual application of it. I also like the word democratization, if I can say it. Democratization of the capability, for me, is one of the great symbols of what AI is becoming. You hear people talking about it on the bus, on the tram, in the trains. I used ChatGPT this morning. It was amazing. For what? But we'll come back to that later.

Sean Corbett: And this is why I think definitions are important. It'd be easy to get onto a nice theoretical discussion, but if you look at the narrative that's going on from the political level down to the person in the pub, the understanding of what it is, it's like this miraculous magical thing that just happens, but it's not. And we need to become comfortable and familiar with both its opportunities but also its limitations to start using effectively.

Harry Kemsley: So let's start digging into that a little bit, Jim, because given that we're talking about this democratization, there is a danger of course that in the open source environment where anybody that can have access to the internet, or other publicly available sources, or even commercially available sources, could start to use these capabilities. And in an organization, like James, where we purport to produce intelligence, valued content from open sources, the interest we've got in AI is obvious in terms of the ability for us to do things at scale, for example, or distract the user, in our case, the analyst from having to do redundant things that humans don't need to do because machines can do better, onto things that humans are still better at, such as the judgment- based analysis. So let's move then this conversation. Now we've agreed there was this capability and we're talking here about the democratization of the capability. What are the uses? Where does the artificial intelligence capability become of use when you're trying to derive intelligence value from open sources?

Jim Clover: So I'm inclined to talk about an application I wrote, but I'm going to stop myself for a second. But in essence, what AI certainly does for me in my open source work is it allows me to analyze and summarize really, really well. So there are all sorts of super clever stuff that goes on in genome therapy, and studying x- rays, and MRI scans, X, Y, and Z. And I love all of that stuff as well. But for me, in that open source Intel text- based world, the ability to take a huge corpus of text from say current events in the last 24 hours, and to be able to summarize that down to one or two paragraphs is something that if I try to do it any other way, the only way I would trust quite frankly is not in code. It's going to be the human to actually make that work. So the ability of a one- man band like myself to, and I will talk about the app because it's easier, is that, for my clients, I download every news article related to inaudible think of some Google search words I say to fetch code. So the thing that goes off to the news agent, so to speak, and says, " I want 20 articles maximum related to technology X," and it'll go around the various news sites that I can get to, that aren't paywalled, and it'll fetch down 20 articles. Once those 20 articles are received, I then pass it to the AI and I say, " Taking these articles of various lengths," I don't know the length of these articles, bear in mind, some could be quite long, " Summarize it down to one paragraph and also extract a headline statement for each article." And then I hand the headline off to another bit of code that puts it at the top of an email, and then it puts the body when you click on that headline. And again, the role of AI in that sense, in that open source intelligence or news- gathering sense, effectively does the work of the analyst that might be summarizing that news article. Now, to Sean's point about risk. Say is it as good as a human doing it? Potentially not. Are my clients that receive these bulletins moaning about the fact that a human could have done a better job? No, I've not had one complaint because they are receiving a bulletin whereby they are still told to click on it for further information. And there is also a caveat at the bottom of the email that says this was generated by an AI, there may be inaccuracies. It's important to make your own assessment. So again, has it enabled that application to exist? 100%. Is it bulletproof? Of course it's not.

Harry Kemsley: Yeah. So summarization and collation, those two steps in an intelligence cycle, definitely sounds like a media application. Sean, so you want to say?

Sean Corbett: Yeah, no, I think there's two really important points there that are sometimes lost. The fact is that you initially defined it parameters. So you said right, within this area, look at this in such detail. So already you've had an intelligent input. Now I don't think everybody does that, but everyone just says find out everything there is about that. So that's where the human loop has come in really early. But from an intelligent perspective saying just try and filter out, and then you've gone the second through level also using the AI goes, okay, now extract the pieces for me. Now both of those things are what we used to do in the traditional intelligence world and you've probably seen me go through Excel spreadsheets, in my inaudible trying to do that ourselves. So really, it's back to AI as a tool to simplify what you want to do.

Harry Kemsley: Yeah, Jim, let's just step round the intelligence cycle, another click round cycle. So we talk about somebody deciding something needs to be done and you've said to the AI, go away and find me these 20 articles and summarize them, give me a heading, etc. So you've gone around the analytical step to some degree, summarization is really what I mean in your case, and you then disseminated that to the community that I've subscribed to your service. That's the cycle in a few words. How much further could the AI get into some of that processing and analytical phase? Could we ask it to do more work in that phase, give deeper analysis, or do we think that's where the AI starts? This is not a leading question, I promise you it's a genuine question, but do you think that's where the AI starts to struggle because it doesn't have the judgment, the experience, that an analyst should bring to it?

Jim Clover: No, that's a really awesome question, Harry. And again, it's one of those areas that when I sit down with people that aren't intermediate or advanced developers and I put myself in the intermediate going on amateur side of life, it's all in the prompt. Bear in mind that we're not talking about somebody typing a prompt into ChatGPT, now we're talking about somebody embedding a prompt within a block of code. So a Python script for example. And yes, 100%, you can go further. So you could say you've fetched the news articles, if you are happy, we'll keep using that example. We've now summarized the paragraphs down and we've got the header out. There's nothing stopping me taking all 20 summarized articles, passing them to another prompt within that block of code to say, your job is the master summarizer. Look at all of these news items, and produce a further overall summary for today's last 24 hours in technology X. So again, it's what's really important, I think, for people. And part of that democratization of AI and democratization of coding is the power of the prompt, is for people to understand that. Because I have not touched the code on that news email bulletin service, I've not touched it since November last year. It runs at half eight every morning. It was 70% written in AI. I just prompted it for the code that I wanted and Alice down the rabbit hole time, I actually got AI to generate its own prompts that were best for the job. So again, it's all in the prompt, bottom line.

Harry Kemsley: Is this what people mean when they're talking about prompt engineering? You hear around the likes of these large language models, if you can engineer your prompts well, effectively you're more likely to get the result you need.

Jim Clover: Absolutely. And then we drift into that a very interesting term that sloshes around, agentic and agents. I wrote something last year when agents were really starting to pop as a popular subject. Again, just wanted to stick my nose in and have a go where I had the researcher deep analyst role and final report writer and had three agents doing that using similar open source research tools. It didn't go well. My wife detected that because this was all running offline and again that might be something else we want to talk about, but she noticed that the graphics cards were screaming in the garage and it was because the agents were actually, not arguing because they're not conscious, but they were in some sort of quality control loop and they were caught in an end endless loop of that's not good enough, go again. That's not good enough, go again. So that was quite fascinating to watch.

Harry Kemsley: The garage was nice and warm I guess with the heat generation.

Jim Clover: Yeah, so don't get me wrong, I've not got a huge data center here, it's just a couple of GPUs, but that's a whole subject in its own right about where I major a lot of my work is in offline AI, rather than using ChatGPT for everything, I still use Anthropic's Claude and ChatGPT, especially for coding, but for summarization and for texture based, I guess you could pass it as alternate work is all offline.

Harry Kemsley: So again, before we go much further, let me just summarize what I think I've heard. This is human summarization, so therefore bound to be inaccurate. We've talked about the fact that the collect, collate, summarize and analytical steps can be, with the appropriate prompt engineering, you can generate actually some pretty high end, high caliber product. I'll come to you first Sean, and I'll come to you second Jim in a second. So what's wrong with that?

Sean Corbett: For me, I don't think we're quite yet there yet in terms of the analytical step. I've been playing with ChatGPT and Harry laughs when I get cross with it, as would you no doubt. But it's been brilliant for prompting. So we've got another podcast coming up. I thought just make sure that my research, I've been covering the right area. So I used it to summarize all the rest of it, but there was one piece that came out, I won't talk about particularly which area but it won't take much to work it out. It started off about all atrocities being committed by one side and it wasn't the question I asked it, so I went back into it and said, " Is this a balanced view? Because there is a different side," blah, blah, blah. Is this validated, the rest of it. And it was ChatGPT, came back and said others are available. And they said, "You're absolutely right. I do need to be balanced and check and do that." And then it produced. I said, " Well just rerun." And it produced almost exactly the same again. So I went back and said, " Are you able to learn from people's inputs?" And I have to say the answer was a critical. So this is why at the moment anyway, there is obviously a certain amount you can trust, but it's back to the old average of garbage in, garbage out. If the resources the reference it's going to are not universal and totally objective, which is where your scripting comes in, your questions, then you're always going to get inaudible. Now that's the same of human beings, of course it is. But a few people that are, these tradecraft inaudible, got to are following the right tradecraft, they are going to be able to get through that and find out what's objective.

Harry Kemsley: Haven't you just exemplified though what we mean by tradecraft in a positive way? You've said yourself, I asked a question, I've got an answer. My experience tells me that doesn't feel like a complete answer, while it may be relevant to my question, but it's not complete, and therefore I want it to be more complete and you've then reprompted it.

Sean Corbett: Yeah.

Harry Kemsley: I think what this talks to, Jim, now coming to you is this sense of unease that people have. And I think there's a cultural issue here as well, by the way, in terms of I don't use technology, I use my Excel spreadsheet, which by the way is in itself a technology that previously was considered modern and advanced when previously had a piece of parchment and a quill and ink. But to get to the point, I think there is a degree of unease because people just don't understand where it's going for the data, what biases might be in the way it's doing that, and how it presents it. Because I saw a phrase just recently that these large language models are desperately trying to please you.

Sean Corbett: Oh, yeah.

Harry Kemsley: And therefore the way you write your question, going back to your point, Jim, about prompt, actually is giving it a clue about how best to answer it. And if you're not careful, you swing it towards a very biased, in quotes, biased answer, it is trying to please you. So this point Jim, and this is a very long way of getting to the question, so apologies. What I believe about the intelligence cycle is that we are pretty comfortable with the idea of collect, collate, summarize, because it's a relatively straightforward process of going out and collecting, collating, etc. What we're less comfortable about is how it then does the analysis and how much bias might be in the AI model or the large language model. And because that's a black box, potentially, we can't see it, we can't go to the source. And I think all three of us will remember moments when on presenting an intelligence product, an answer to a question, we've been challenged. I said, " Well how do you know that? Show me your sources, show me analytics, show me the hypotheses you came up with and how you just proved the other ones." That's the bit that I think perhaps gets in the way of cultural and that practical application of some of these technologies in that analytical step.

Jim Clover: Yeah, I mean I think it'd be useful to carry on the previous example for the audience as well as for this discussion. So what I'm asking in my application is for a... And that leg of things is actually not AI. So requesting out to news engines using a paid for service that I use, I'm literally sending an API query. So I'm sending out a server query and saying, " I'm looking for stuff on quantum computing UK." And I send some parameters in the API call saying for the last 24 hours. So the returns I get are a load of web links and those web links come in. And because the next stage, again, is not AI, it is go and fetch the content of those news articles if you can. Some of them won't allow it and so there'll be some retries, that's not AI. But what I end up with is blocks of text for each news article whereby I'm not asking the AI for review, as in Sean's example, I'm telling it to summarize based on what it's being handed, which is those independent news articles. And I'm very, very strict in the prompt. And again, we're back to prompt foo. So in the prompt I'll be saying do not use your own knowledge to summarize this article. Do not enrich it, just provide an accurate summary of these many paragraphs if they exist into one crisp paragraph. And I think it is a very strange thing, and I don't think it's purely a British thing. I've noticed that people don't tend to get too stroppy with prompts when they should. And I think what I have found is the more explicit I am with prompting, especially in the engineering side, I did a vision model, which is probably inaudible where one of the vision language models where I passed a picture of a damaged bridge for a client that I was working with and we just had a discussion over a cup of coffee and I said, " You know what? I'm going to go and try this. It's got to be worth a go." And I passed an image of a damaged bridge and said, " Show me where there is paint peeling, rebar exposed," and it went off on a right tangent. There was planes landing on the bridge and cats and well, not anything that outrageous, but bottom line is it was highly inaccurate. And what I realized was that I had to break it into sections and treat the LLM like a child almost and say, " Is there any paint peeling? Only answer yes or no, do not say anything else but yes or no." And it got it right. Again, but I had to reset in between each one. I couldn't allow it to remember its last context for example, because that might sway. So again, these little prompt techniques, yes, we're getting into more code side of life, but they actually work for people that are just using ChatGPT or other models apply, is that this really does come down to getting strict with those prompts, and-

Harry Kemsley: Which does actually start a point, Jim, doesn't it, to the need for, dare I say, human in the loop. You need to be more engaged. You can't just let-

Jim Clover: Absolutely.

Harry Kemsley: We've heard from both Martin and Keith previously, and I think as well from Harry about hallucinations, which I'm sure inaudible are aware is a real problem. Something else I heard about recently, by the way, from a data scientist who specializes in AI in NATO, that some of these AI models are now suffering from what he described as dementia, that they're actually becoming degraded because so much of the content is self- generated, it gets itself into a loop where it starts to degrade its own perspective. This is using your example of the cats arriving on the bridge. If it introduces that into its own understanding of something, then it will continue to use that, unless it is told to not do so, which suggests that the need for a human to govern it remains. It's not yet got to the point where it can do this for itself. So does that point towards a moment in the process where humans must be involved in the intelligence cycle?

Jim Clover: Right now, and it goes back to Sean's point, which I agree with is that it's not ready for prime time in that sense. But let's just pause for a second because I think there's a risk of sounding too negative on AI as well in the sense that what's really important is that we've got a technology we can fall out with in this way. We've never been here before where we can actually, instead of endlessly searching Google for the answers, or phoning around as many specialists for review, the fact that we can actually have a discussion on vast waves of information that are locked up in these huge LLMs, and the human in the loop is crucial because ultimately, and I'm sure OpenAI and Anthropic and Google an Gemini and all these other companies would say, " It's up to you at the end of the day whether you accept the answer or not, we're just giving you access to a corpus of inflation, over billions of books, manuals, code, whatever it is." For me, the sort of summarization of AI for me in 2025, is it's a productivity tool-

Sean Corbett: Yeah, I can see that.

Jim Clover: ... for thehuman, but there have been some more recent, I'd like to call it emergent technology because all AI is emergent. It's not new per se. There's things like the model context protocol work that's just coming out. And again, the ability for models to now branch out from just relying on their knowledge to be able to use these model context protocols. Just think of them as small applications that really know where other stuff lives. So you can ask the question of the AI and say, " What does my company know about X?" As long as your developers have plugged in via the MCP, the model context protocol, the ability to talk to company data, then again you are going to get a far more accurate, non- hallucinating response. I think that's really exciting. I think MCP, model context protocol, is emerging as something that could really help drive accuracy and utility, but for me it's still human productivity going 10x.

Sean Corbett: I get it. It got my thoughts going to a word that we use quite often, we've covered, and I know Jim, you've done some work on this, but ethics. People talk about ethical AI, well that's giving it some sort of soul. The ethics is the person to put on there. And we've talked about many times about ethics generically in no simple terms of that which is right, which doesn't necessarily mean it's legal, the rest of it. And I know you did some work with putting in some things, create a bomb, that sort of thing, and you're not allowed to do that.

Jim Clover: That's right.

Sean Corbett: But surely that's due to those parameters are what's defined within the programs anyway, so it's still back to human.

Harry Kemsley: Okay, we'll take just a short pause there. That's the end of part one. Please do join us for part two very soon. And thank you for listening.

Speaker 1: Thanks for joining us this week on the World of Intelligence. Make sure to visit our website, janes. com/ podcast, where you can subscribe to the show on Apple Podcasts, Spotify, or Google Podcasts so you'll never miss an episode.