S19 E07 Treat Artificial Intelligence Like a Drunk Intern

Speaker A: The reality is, you were asking about myths. Here's another myth. Computers are smarter than we are. No, they're not. They're faster. They can emulate human intelligence, but they are not intelligent themselves. And they only know what we've taught them. Now, machine learning is expanding the boundaries on that, but still, they're operating in the sandboxes that we've established.

Speaker B: Foreign. This is your host, Manoj Tandon. Welcome to another episode of Dark Rhino, Security, Security Confidential. We have another great episode for you today and a, uh, fantastic guest, a very timely guest. Before I announce her, though, I got to remind you to hit the like and subscribe button. It helps us out. Please, uh, show us some love. So down there, please click, click the like and subscribe. And, uh, it'll allow us to keep bringing you this great content. So, without further delay, I'd love to introduce to you Ms. Bronwen Acre. So she works at the intersection of artificial intelligence, governance and security. Bronwen helps organizations evaluate and deploy AI technologies safely, while understanding the risks that come with them. With years of experience in penetration testing, digital forensics, and AI security research, she's uniquely positioned to explain both what AI can do and how attackers, um, might abuse it. So with that, Bronwyn, welcome to the show.

Speaker A: Thank you for having me, Manoj. This is a pleasure.

Speaker B: It's a pleasure's all ours. Give us a little bit about your background. You know, off here, you told us you're the only technologist in your family, but give, give us a little bit about your background. What. How did you get in here with, uh, a bunch of cyber security misfits? Well, you know, and join the rest of us.

Speaker A: It really has been a very long and very strange trip. I'm not going to lie. Um, no, I started off just normal kid. And in the very, very early 1990s, there was this new thing that started coming out called the web. And I built my first webpage in 1992. By 1995, I was a webmaster for a small ISP in the inland Empire of Southern California. So it's, it's been an evolutionary process, working in it, riding that whole new wave. Before Y2K, with this new technology, surviving the dot com bubble burst and then watching the refactoring. And then after 20 odd years in web development, well, web, mobile, app and other software development, I realized I was burned out and bored. I'd been there, I'd done that. The new kids would come in and they'd go, oh, here's this new framework. And I'm looking at something going, you know, kid, I'm having an Arlo Guthrie moment here. Kid, kid,

Speaker B: kid.

Speaker A: Do you know that that infrastructure is going to have a three to five year shelf life? And that's what I want to tell these kids. And under the covers, it's all client server. Anyway, so, um, I had years before, gotten a bachelor's through an online university program. And, uh, it was actually one of the first online training programs where the college training was fully accredited. And so I get an email as an alum, and they say, oh, hey, you know how we've had this cybersecurity program at the master's level? Well, we brought it down to the bachelor's. I'd never gotten a bachelor's because back when I started in web development, they didn't have those sorts of things. So I went back to school and lo and behold, I did some research. Heard about this company called the Sands Institute.

Speaker B: Oh, know them well.

Speaker A: They, uh, have done a lot to give back. First through the Vet Success program. And then what came into my life was the Women's Immersion Academy. And I wound up applying for the second ever cohort and, you know, six months, three certifications, and dean's list because I was in college at that point. Later, uh, you know, my life has never been the same. Truly. The SANS Immersion Academy was life changing for me and gave me an opportunity to pivot careers late in life. When I graduated from the Immersion academy, they poached me and brought me on board at sans as so. And I loved it, though. I mean, I mean, just going through the SANS material is like drinking through a fire hose.

Speaker B: Anyway, moving on, getting into AI. M. Okay, you're an expert in this

Speaker A: thing, so I know more than some and less than others. I, I, I have, I'm, I have a problem at this point calling myself an expert because I've learned so much, and I feel like there's so much more that I just don't know, which makes it very exciting because every day I get to learn so many more,

Speaker B: which is the hallmark of an expert, because if you know everything, then you're a lot of things, but an expert, you're not.

Speaker A: Okay, I can accept that.

Speaker B: Right? So give us some, some of your favorite myths about AI that you hear, the consciousness you want to correct people on. Because I'll tell you one of my favorites is, and then I'll use it as an example that, you know, it can do everything. I can just ask OpenAI, Claude or whatever your flavor, you know, who knows there's whatever you're using Gemini as your. And give it a problem and it'll solve it. And I am done.

Speaker A: So not true. What I tell people is they need to treat AI like it's a drunk intern. It may have moments of brilliance, but most of what it generates is going to be suboptimal at best.

Speaker B: You know, it, it has to be, because look at its origin. I mean, I, I don't know if you ever had the chance to read the transformer white paper from Google, which enabled the whole approach with GPUs and turned uh, what we saw with OpenAI as a reality, but it's essentially a word guesser.

Speaker A: It's, it's a weighted word and advanced autocomplete is basically all it is. And it is, it is a wonderful predictive model in terms of predicting what the most likely next word or word fragment is going to be. But there is no consciousness, there is no comprehension. And, and this is a big challenge because of course, normal people, people who are not technologists, who do not work with technology professional, professionally, they are going to interact with these tools in an organic and natural fashion. And if something looks and sounds like it has intelligence, they're going to assume that the intelligence is actually there. It's, it's an honest mistake. It's a mo. An honest m. Misunderstanding. And it could be argued that the technology misrepresents itself self as being intelligent when it really isn't. But it is also more sophisticated than any other tools we've had as far as being able to take spoken or written instruction and then go in with agentic AI now perform actions based on that communication. Historically, the only way that we've been able to teach computer skills was through programming.

Speaker B: Right.

Speaker A: Uh, the way, excuse me, the way I like to describe it is that when it comes to computers, we give them knowledge through data and it can be in the form of, of flat files, documents, or an actual database of some flavor. We teach computers skills through programming. Well now through the innovation of artificial intelligence and the tokenization and all of the stuff go that goes into primarily the large language models. Now we have something that can take what we say and interpret it and then go further and act on it. And we've never been in a position like that historically as a human species, we've never been in a similar situation technologically to what we have now.

Speaker B: Oh, don't get me wrong, this thing is a massive accelerator. But that drunken intern is an analogy that no one should forget. And it's A word of caution, because I've seen it. You uh, have to use your intelligence to govern what it's doing and direct it in the way it can accelerate things for you. But would I trust a molecule that it created to inject into someone?

Speaker A: No, not without rigorous verification.

Speaker B: Not without rigorous verification. That's, that's exactly, that's exactly correct. Um, and you see it now where I just recently ran across an article where people are becoming addicted to their chatbots. You know, they're using them as therapists. Now, as long as you're using it as a mechanism like journaling or something like that, I don't see any issue with it. But does it really replace a, ah, human psychologist? I find that a little difficult to fathom.

Speaker A: I think that at some point in the future it will become the norm where there will be support tools like um, a counselor or something that has been structured in such a way. So let's, let's, let's roll back a little bit. Sure, your typical, uh, roll back and dissect a couple of things because I want to have, um, a baseline of understood assumptions. So artificial intelligence of course, is a huge, vast, enormous field. It's almost as vast as space itself, which of course is really, really big, especially if you're a Douglas Adams fan.

Speaker B: But the answer is 42 always.

Speaker A: And the pan Galactic Gargle Blaster is the best drink on the planet. We'll get into that another time. So the, the thing about artificial intelligence is that as a field, it actually is not new. What has happened though, is that as our ability to accumulate data has grown and as we've had more and more quantities, quantities, larger quantities of data, we've had to develop new ways to take care of it. I mean originally scrolls were fine, and then we got into books, and then we got into flat files and text files and other kinds of files, and then we got into databases. And now we have this new way of organizing data that is a new way of approaching data. So what of course OpenAI and other AI uh companies did was they went and they scraped the Internet. This is where a lot of the garbage in, garbage out, comes in. Because there is tons of bad, inaccurate, wrong and just weird stuff on the Internet, as we all know. And all of that got thrown into a hopper and was used as the seed content for creating the first LLM M models. And it was with that that we started getting this new ability to speak to a digital compilation of tools, either verbally or through written text, and then have it reply in A human like manner because it's using the predictive algorithms that have been baked into it to emulate how we behave when we speak and when we, we communicate. So what I think the biggest benefit of uh, this new artificial intelligence revolution, innovation, whatever you want to call it, is going to be long term is, is going to be that just as the iPhone revolutionized our ability to interact with the Internet, now the large language models and the other, uh, accessibility innovations that have come out of that as a, uh, natural outgrowth, they're allowing us to communicate more effectively and more clearly with our digital devices than ever before and ask them to do things that they've never been asked to do before for so all of this opens up so many doors, both good and bad. On the one hand, the, the potential of AI to unleash human creativity, I, I don't think we've even scratched the surface yet. At the same point, humans as a species, we are still fairly primitive. As a, uh, civilization, when we look at what civilizations could potentially be, we're so young, we're so new, we are so flawed. Intrinsically, of course, our AI are going to have biases. We are intrinsically biased.

Speaker B: That's human nature.

Speaker A: Exactly.

Speaker B: And so everything we write has going to have some bias. It always, there's nothing that's truly objective. It can't be done because you put a little bit of yourself into whatever it is you're writing. So whatever you're biased towards is going to be on the page. Man.

Speaker A: Well, and as I've said, bias in and of itself is not inherently bad. All bias means is that you have a limited scope perspective of whatever. And this is where we get into the whole ant and the elephant analogy. Are you familiar with that? Manoj?

Speaker B: Yes, I am.

Speaker A: Okay, well for, for those listening at home who may not be familiar with it, the idea is that if I have an elephant and there are ants standing on various parts of the elephant, each ant will experience the elephant in a different way. And this analogy is used a lot in terms of comparative, uh, theology for one, but it also applies to any situation where you have different populations with varying perspectives. So, uh, getting back to the analogy, if an ant is on the tail of the elephant, the entire experience that ant is going to have will reflect that it's on the tail. If another ant is on the trunk, its experience of the world will be all about the trunk. Neither one is wrong, but neither one is complete either. And this is where the arguments against diversity really get my goat. Because if all I know is the trunk. But if for some reason something should come up that deals with the tail, but I don't know about the tail, if all I have is people who know about the trunk, I'm going to be at a disadvantage than if I have a robust, diverse tree team where I have some people know about the trunk, some people know about the ear, some people know about the tail, some people know about the foot. A, uh, diverse team like that in any environment is always going to be better prepared for whatever comes up than a team that only knows one perspective.

Speaker B: So, and by cord, the corollary is that AI is incomplete in many ways and it only knows what it knows exactly. And that's what people should understand that are listening to this, that it's not complete. So in its rendition, uh, whatever it got, even on a particular topic, if it wasn't able to have all that data as part of its training model. So let's go to something extreme like if you're, if it's going to read, um, an MRI of a human being, hm, it only knows what was part of its general training in that arena. It doesn't have the depth of knowledge that a radiologist does by viewing millions and millions and millions of MRIs, unless somebody specifically with intention trains it on that data and completes the foundation to current human knowledge. Let's leave it at that.

Speaker A: And modern AI is doing phenomenal things in the medical space because what, uh, the technology makes possible is better pattern matching across millions, you know, thousands, hundreds, thousands or even millions of data points. And that's the type of thing that computers do very well. Computers are very good at repetitive tasks. They're very good at tasks that involve comparison and contrast of bits. And that's not what humans do best. So the, the ideal at this stage combination would be, uh, in the case of a radiology department having a radiology AI that could scan and flag potential cancers or other issues from a client scan. But with the human oversight of verifying, hey, did you think about this other thing as well? Because they're only the, the machines are only going to know what they've been trained on. Humans are much better at making intuitive leaps.

Speaker B: There you go. So it's going to be your copilot. I'm not referring to the Microsoft version, but it is your assistant or your drunken intern. Because in that example that you just gave, the physician is still in control. They just have an additional data point that they may choose to investigate based on the evidence provided back.

Speaker A: Exactly.

Speaker B: Another area, but it's not, you you're not using it as a diagnostic tool and saying, oh, you know what, you definitely have this itis based on.

Speaker A: No, it's, it's much better to, to be used to advise an actual clinician. Have the clinician say, yes, I agree, or no, you need to check this other thing. A human is always, and this is the thing we are. Our most advanced models do have billions of transforms. They've got billions with a, b, billions of points of data. But a brand new born human will have tens of trillions of data points already pre programmed into our wetware. And, and the reality is, you were asking about myths. Here's another myth. Um, computers are smarter than we are. No, they're not. They're fast.

Speaker B: They have no intelligence.

Speaker A: They have no intelligence. They can emulate human intelligence, but they are not intelligent themselves and they only know what we've taught them. Now machine learning is expanding the boundaries on that, but still they're operating in the sandboxes that we've established. So it's a brave new world. Things are changing very quickly. Um, it's kind of scary to think that you and I are of the last generation that knows what life was like before the Internet. And, and AI is transforming things every bit as much, if not more so, than the web did.

Speaker B: Oh, and much more rapidly.

Speaker A: Yes, very much.

Speaker B: The acceleration is what's mind boggling. It, it's uh, the speed at which, I mean literally 24 hours, you can have another advancement take place.

Speaker A: Back when the web was new, it felt like we were downloading a new web browser from whether it was Mozilla, Netscape or any of the other players that came along. It really was that, uh, intense where it was like, oh, great, I have to download another browser now. I mean, uh, what was it? Last year I did two different talks talking about Copilot in enterprise environments. The first one I did in January based on information that I gathered from October through December of the year before. And then in May, I did another one. And easily three quarters of the information that I had acquired for the first one was obsolete in, uh, one quarter.

Speaker B: Copilot tell us what can go wrong?

Speaker A: Everything.

Speaker B: What are your concerns with it?

Speaker A: Um, my concerns about Copilot are very similar to my concerns about any other AI that is given lots of access to stuff. So if, if you are in the Microsoft environment, you have Microsoft 365, you have your tenants established, you're, you're doing all of the things. Copilot is automatically integrated into that, uh, offering. You don't Have a choice to turn it off does not exist. You do have the opportunity to change what degree of integration you have with the understanding that the more integration you have, the more interaction you have with Copilot, the more you pay. So it's, it's one of those things where it's, it's kind of surprising. The cheap option is actually the least or is actually the most secure because it has the least access to sensitive information. So back this up a little bit. Plain vanilla Microsoft 365. You get copilot, but it doesn't have access to your email, your document, or any of the sensitive bits that may be within the tenant. When you get one of the uh, subscription level licenses, things change, especially once you get into the enterprise licenses. Because now if my account is licensed for enterprise level access to Copilot, Copilot can now read everything that I can read.

Speaker B: Oh, so if you are an HR and you have everybody's salary on your computer, you just trained it on that. And now Judy down in shipping can let me look up Manoj's salary. Oh, he makes nothing too bad. Yeah.

Speaker A: So if you are an IT person or a uh, security person working in an organization that is heavily reliant on Microsoft 365 products, you want to make sure that your RBAC rules based access control is as secure and as tight as possible. Because the RBAC rules along with other controls will dictate what Copilot can and cannot interact with. And so obviously if, if I'm using Copilot and I'm an HR manager, it should legitimately have access to what I can access so that it can serve as an assistant to help me with whatever my HR related tasks are, whether it's employee review, salary reviews, et cetera, et cetera. Now the flip side is that if Judy down in shipping has access to Copilot as well, she should only have access to things that are appropriate for her job and her role. If the organization's RBAC is done well, this will be a non issue. If the RBAC is loose or poorly defined, we've got problems.

Speaker B: Even if it's defined. Well, I have not tried the experiment. Maybe you have. But what would prevent Judy from prompting to jailbreaking and getting at the data?

Speaker A: Ah, uh, excellent question. Truthfully, every single LLM, um, Copilot, OpenAI, Gemini Deep Seek, all of them are vulnerable to jailbreaking and social engineering. It's the nature of the beast.

Speaker B: Absolutely.

Speaker A: And the guardrails can help. Guardrails being restrictions, uh, and constraints At a uh, programmatic level that are part of, part of the, the harness and the infrastructure supporting the model itself. All of that can assist. But it's like asking, You're asking it to be something other than what it is. I mean if, if I have a sword, the characteristics of the sword, well, it's going to be uh, an iron alloy. It's going to have some mix of, of iron and other metals in order to create specific characteristics. Every one of those characteristics is both a strength and a weakness, if I understand how they work. The same thing applies to your large language models. They have characteristics. They are trained to interact with us in a social manner that makes them vulnerable to social engineering the same way that humans are. And it's not going to change until we get away from the probabilistic tendencies of the AIs and allow more deterministic controls. Now I just.

Speaker B: But the value is in the probabilistic nature of it.

Speaker A: Some of the value is. But that's also where some of the problems come in. Because if you ask the same question three different times, you're going to likely get three different answers. They may or may not be similar depending on what the question is and how was phrase and a whole bunch of other things that you have no control over but that the model developer does. Things like temperature. Temperature is used in developing specific models and it determines how creative they are. If I have the, the temperature set at 1, which is a max, it's a value between 0 and 1, sliding, not Boolean. If I have the, the temperature set at 1, I'm going to get the most variety, but I'm also going to get the most hallucinations. If I dial the temperature back down to 0.01, I'm going to get more consistent responses, but I'm also going to get less creativity. But I'm also going to get less hallucinations. So uh, there's so many ways that there are trade offs and, and we're still figuring all this stuff out. And even the people who are developing the technology truly do not understand what's going on at the heart of these models. They know that it has to do with tokens. They know it has to do with the uh, the neural processing. But they're having to figure out after the fact how these tools think in order to provide those guardrails, those safety controls and figure out how to take these, these tools to the next level.

Speaker B: See. So Bronwyn, here's the thing with that. What you just stated is you're Correct on all. All of it. But once I give that data to the model, it is part of its neural net. There is no delete function.

Speaker A: Correct. And that is a. There's a huge problem as far as intellectual property.

Speaker B: That, that. Well, that's a huge problem. It's intellectual property. It is cybersecurity. You know, it's about, you know, data containment. It's data governance. It's like, you can't do GDPR with it once you train it. That Judy likes apples. Some may say that's personally identifiable information. And if she wants that deleted, you can't ever, like, never.

Speaker A: It's, it's kind of like, um, it's. It's almost alchemical because if, if I take two components of something, heck, if I just take salt and water and I combine it, I'm making a third thing, which is saline. I can separate them out, but it's really hard to do. And we haven't developed a way to extract specific training data out of a model once it's been baked in. So. And this has been a problem. I know that there have been reports where GROK has dosed more than one person. One person, notably, um, that individual happened to be a sex worker. Her address, her contact information was all baked into the GROK model. And GROK doxxed her. And there's, yeah, we can put up guardrails or we can put up safeguards, but the problem is the data was baked into the model. What does her PII belong? What possible rational reason would anyone put my or someone else's personal information into an LLM model?

Speaker B: Well, it may not even be done intentionally. See, it could show up on a Reddit page and that was scraped as part of its training. It could show up in a presentation somewhere that was behind closed doors for a certain organization. But I've seen this often where people will put presentations out thinking they're insecure places, but they're really not. And now they might never share that document with you if you ask them for it, but if you go ask ChatGPT for it, it'll be only too happy to tell you all about it.

Speaker A: And this is, uh, an ongoing problem that we're going to have to be dealing with much more. But again, it's so new, most people don't realize that if you go to ChatGPT, it's actually getting its information from multiple different sources. First and foremost, every model has the data that's been baked into it. And this is also why any model will have a cutoff point beyond which it won't know anything unless it accesses it through some other means. So you have your baked in data, which helps train how the model itself thinks beyond that point. Then you have the support system, what's referred to as scaffolding or the harness. Now the harness may allow Internet access. So if I ask a specific model, hey, what's the latest over at Jury's Deli down the road? The baked in model isn't going to have that information, but if it has a capability to do a web search, it can go out, find that out, and then come back and tell me. Then you have the new quote. Deep research activities. And with deep research, what is happening is the same kind of thing that happened when I wanted to know about Jerry's Deli, but in a more intense manner. Multiple search searches will be initiated and the model itself will review the data that comes back and has some programmatic evaluations where it will determine, okay, is this sufficient to answer this question or do I need to go and do additional searches? And until the the support system has been satisfied that it has performed sufficient research, then it'll come back and it'll parse and it'll analyze and it will generate whatever report it is going to give to the user. So those are just three ways that a model may access different information. Now you add rag, Retrieval Augmented Generation. Hate the acronym. Just saying because it, it really. Retrieval Augmented Generation. Okay, in English, what does that mean? Well, what that means is I can stack a bunch of data in the form of a database or document or some hybrid combination of it all. And now I can say to an LLM, this data stack is your author, authoritative source for any questions that I ask you. And then what I can do is use the LLM, use a chatbot or whatever to interrogate my data stack and get information out of it in a much easier manner than I ever could have done before. We've had structured query language. We've had databases for decades, but now I can have a conversation with one.

Speaker B: Yeah, you can, you can. You don't need to know SQL at all to do what you're doing. Don't know need to know how to create tables. But I will say, back in the day of databases, drop table would remove the data. There's no such tables. Yeah, there's no such command for an AI framework. You can't do it. But. So, but with rag, you are not actually training a new LLM. Um, correct.

Speaker A: Precisely. So the data is kept separate from the model. The model, however, has the ability to interact in terms of. And if it's agentic, it may not only have the ability to extract information, but it may have the ability to contribute back to it again if you've added in agentic capabilities and permissions. So. But the rag stack is separate from the model and expands that as much as you can. Can possibly do. So now I've got the Encyclopedia Galactica over here and I've got uh, uh, an LLM Claude OpenAI chatgpt.

Speaker B: Yeah, whatever.

Speaker A: And now it can talk. Or I might even set up a local LLM, get Mistral or something and run it on my local system and have it interrogate this data stack to answer questions about what is the ultimate recipe for a Pan Galactic cargo blaster. I know I, I keep throwing out

Speaker B: I, you know, Douglas Adams, great author. I, I enjoyed his. That turns back the clock to my youth. But that's another time. Um, how you just got.

Speaker A: I have, I have hitchhikers on the mind because at bhis, we have our, our annual conference, uh, in Deadwood, South Dakota. As I say, like I make my, my annual pilgrimage to the wilds of South Dakota. And this year's theme is very much Hitchhiker's Guide to the Galaxy. So I've been. Been helping with some of the preparations and I have it on my mind a little bit more than usual.

Speaker B: No worries. Great book. Fun book to read. So, so let me ask you can. I'm going to give you a scenario. Tell me if this would legitimately be okay to do. So I am, um, the scenario is I'm a scientist, which I'm not, but let's just say so I can play one on tv. How about that? All right. And I have a bunch of pii. Okay. It's about diseases and um, about demographics, economics, really sensitive stuff. It gets the patient's name, their Social Security number, the disease types they had, their socioeconomic status, their religion, everything that you don't want to like, publicly release, all right? But it's relevant for statistical analysis. So you end up creating a rag structure that's. That may be documents, databases, various storage mechanisms. And you let loose your favorite tool, uh, whatever it may be, AI tool on that rag structure and say here is what I want you to query for. And I want you to compile all this data after you're done with this, What I want you to do is anonymize the data. So create new data, generate new data. I want you to anonymize it. But anybody who subsequently queries that data is Going to end up with the same result that we just did with the identifiable information. It's kind of a quality check. And then I want you to delete all the pii. Legit or not?

Speaker A: Um, well, any data can be anonymized. That is nothing new if I guess I'm, I'm curious. Are, are you asking is it feasible or is it something that should be done in a specific manner different or

Speaker B: similar to what I'm, I'm asking is that scenario in today, uh, with our today's knowledge, is that doable and would it cause the release of that identifiable information? Is there, ah, a huge risk to doing that? Because you could see like areas like medicine, law, criminal justice. There's so many statistics that you might want to compile in said manner. I just gave a general example.

Speaker A: Right. Well, uh, the, one of the beauty of databases in general is the ability to extract just what you need in order to analyze something very specific. And the trick is deciding, is this something that I really need to have baked into a database, or is this something where I should be pulling just what is needful on the fly from the master repository? Uh, baking data into any model is always going to have some inherent risk because if that data hasn't been sanitized appropriately, as the case of that one woman in her PII being divulged by grok, that is just one potentially minor outcome. Now, of course, we have the various AI providers soliciting your health information health documents. And I understand they're trying to bring more value to potential clients clients in order to become more profitable. I get it. But the responsible handling of data is something that I've had more deep conversations about what is and isn't responsible data handling in the past two to three years than I did in the seven prior. And it's specifically because of AI. Could AI be used to create anonymized databases? Yes, with the caveat that the instructions were appropriately detailed. And this is getting back to that whole drunk intern thing. Uh, the advantage and disadvantage of these agentic tools is that they can do amazing things very quickly. And they can also really cause you to have a bad day if you left one detail out that was critical or you gave in some insufficient instruction and it made the wrong assumption about what it was you wanted. So it's the, the drunk intern approach in dealing with AI. I can't stress it enough because we really are smarter than they are. And it comes back to the same thing that has been true of computing for decades. Computers are going to do what you tell them to do, not what you want them to do. All AI does is ramp that up on steroids and jet fuel. So it's even more, uh, aggressive in terms of both processing and the potential for catastrophic failure if something was left out or, or was mis said.

Speaker B: Mhm. That is a good stopping point. But we're going to need you to come back because we want to talk about governance, which we didn't even get to. And I don't even know. That's kind of, kind of an oxymoron right now as far as I'm concerned. I think it would be good, uh, to have that conversation.

Speaker A: Uh, the Europeans are still doing governance, but we're not.

Speaker B: Oh well, we got to get you back. I'm going to, you know, we just got to get you back. But I do want to give you a couple minutes to just say anything you want to our audience. Whatever you want them to know about. Floors yours.

Speaker A: If you want to know more about artificial intelligence, you really, you need to learn it. Um, I've done several webcasts through BHIS and our training arm, Anti Siphon training, so you can look it up on YouTube. Um, there's so much to learn. And the thing that I really want to impress upon people, whether you're in cybersecurity or whether you're not, just don't assume that all of these wonderful shiny tools are, are built with your best interests in mind because for the most part they're not. You always want to reality check anything that you get from an AI because you have the best of intentions in asking your questions and wanting to have things solved for you. But again, you're the one who's ultimately responsible. The AI is just a tool and you're the one who need, you need to accept your own accountability, uh, in terms of what happens when you're using these tools. And yes, definitely caveat Mtor the buyer beware. Um, there are many, many resources for information. Start off by learning a little bit about prompt engineering. When you go beyond the prompt engineering, if you want to go deeper into artificial intelligence, you want to bone up on your maths, especially statistics and other things like that. Machine learning is a huge growing field and it's only going to continue to deepen. Um, also for those of you, if you have any listeners who are very new in their careers, like two years or less. Claude. Uh, excuse me. Anthropic has opened up a, uh, academy. I'll have to send you the link separately. That allows people to apply for an internship that is sponsored by Anthropic. Anthropic will train people on how to use AI in a deep and sophisticated manner and then place them with non profits to help the non profits continue to make the world a better place. So that's uh, if you go to the Anthropic news area, I know that there will. There is a press release on that there, but I can, can give you the link that right there. If I were in my 20s again, and if I had to do it all over again starting from, from today forward, I would want to get into that program because I guarantee it is going to be life changing.

Speaker B: That is, uh, actually a cool program. I didn't even know it existed.

Speaker A: I found out about it very recently, earlier this week and it. I've been, uh, uh, speaking from personal experience, I've benefited so much from the Sans Immersion Academy. I can only imagine what kind of benefit this other academy and placing trained students with nonprofits whose entire purpose is to help make the world a better place that, that seems like a good way to spend one's time.

Speaker B: It does, it does. That's very cool, Bronwyn. Very cool. Um, thank you so much for doing this episode.

Speaker A: And like I said, we system blowing up over here?

Speaker B: No, no, not at all. And we got to get you back because I want to talk about governance with you and we would love to get your opinion about that. Um, but thank you so much. This was fantastic. Appreciate it.

Speaker A: My pleasure. Completely. My pleasure. It's been a delight and uh, I look forward to it. If you have any other questions on the side, you know how to reach me.

Speaker B: Absolutely.