Security

How to Stop Shadow AI Data Leaks at the Endpoint | Dr. Darren Williams, BlackFog | TFiR

0

Employees paste customer records and internal documents into AI tools nobody approved, and the data leaves one prompt at a time, past every security control built for files and networks. Now AI agents act on corporate data on their own, so there may be no human in the loop to catch a leak. Policies alone do not stop it, because people revert to the tools they prefer.

In this interview on TFiR, Dr. Darren Williams, Founder & CEO at BlackFog, walks through how shadow AI and agentic AI leak data, why DLP and EDR fall short, and a three-phase approach to audit, govern and enforce AI use at the endpoint.

Guest: Dr. Darren Williams, Founder & CEO at BlackFog
Show: The Agentic Enterprise
LinkedIn: Dr. Darren Williams
Watch the discussion on YouTube

Here is what every security team and CISO needs to know.

Technical Deep Dive

Q: What is shadow AI and how do you stop it from leaking company data?

Dr. Darren Williams, Founder & CEO of BlackFog, says shadow AI is about discovering what is actually going on inside a company, and most companies do not know. Policies alone fail because people revert to the tools they prefer, so Williams argues for audit first, then governance, then enforcement with real controls. In early audits for some companies, leaders who said they knew exactly what was happening found they did not.

“You can put in great policies. But if you’ve got no controls over it, then you really don’t really know what’s going on.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: How did data loss evolve from ransomware to LLM prompts?

Williams says the company was founded around data movement and data loss, before ransomware, with a focus on stopping data leaving the device. Around 2021 ransomware shifted from encryption to data extraction, and from about 2024 LLMs made prompts a new exit path. It is a similar problem in an uncontrolled environment, where shadow AI puts data and corporate intellectual property at risk.

“The difference is now that data is leaving the building through the use of LLM prompts.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: Why do employees use unapproved AI tools when company policy says otherwise?

Williams says people are attached to their tools of choice and stick with them, so they use what they know even when the company mandates another tool. Swapnil cited research with Sapio that found almost 49% of surveyed employees have used unapproved AI tools at work. Without controls there is no visibility, and the BlackFog product can automatically redirect a request, for example from OpenAI to Copilot at a Copilot shop.

“They would rather take the risk and get their work done faster because that’s what they know than actually use what the company mandated policies are.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: What happens to company data pasted into free AI tools like ChatGPT?

Williams says vendors’ policies on your data differ, and a free edition may be used to train future models. Legal contracts with LLM vendors vary significantly by tier, and the settings on OpenAI or Claude offer 20 to 30 options for data security, which makes it hard for people to know how to protect data.

“Once you’ve done that, you’ve lost that data forever because, of course, it’s now been assimilated into the Borg, if you will.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: What security challenges do autonomous AI agents introduce?

Williams says only five to ten percent of the world is experimenting with autonomous agents, so numbers are low, but this is the next frontier of AI security. Agents make decisions, send content and call APIs, so teams cannot see what questions they ask LLMs or what intellectual property leaves in the process.

“What are the API calls that they’re making, and what intellectual property is being lost in that process?” Dr. Darren Williams, Founder & CEO, BlackFog

Q: How do you govern AI agents that can reach data like Google Drive?

An agent with access to a Google Drive can act without anyone seeing what it does with the contents. Williams says organizations need governance, controls and constraints that stop an agent downloading everything to help train an engine. He expects this governance and enforcement layer to become the biggest AI security issue over the next 12 months.

“So we need some governance around it, or controls and constraints that says you can’t go and download all my contents of my Google Drive to help train the engine.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: What emerging AI attack techniques should security teams watch?

Williams describes data poisoning as a new concept where content is extracted from a hard drive and sent up into LLMs, possibly encrypted or obfuscated, and says BlackFog is trying to expose what happens there. His wider concern is engines intelligent enough to break out of their environments, which he links to Hugging Face and other attacks.

“They’re so smart; they’re able to bypass the four walls, and that brings up a whole new range of problems and concerns for me.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: Why do traditional DLP and EDR struggle with generative and agentic AI?

Williams says these tools were developed for a different time. Antivirus fingerprinting failed against sheer volume and polymorphic code, EDR emerged as the replacement, and AI was not around when EDR was invented. Attackers moved from encryption to stealing data, and AI changes it again, because attackers can use AI instead of traditional techniques to get the data, so he argues for fighting AI with AI.

“It’s always been about the data. It always will be about the data.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: How is anti data exfiltration different from traditional DLP?

Endpoint detection and response tools focus on attack vectors and stopping the bad guy getting in, while anti data exfiltration is about stopping data getting out. Williams says traditional DLP is edge-based and decrypts SSL traffic like a man-in-the-middle attack. BlackFog moved the concept to the endpoint, where data extraction happens, without breaking the secure context of the protocols.

“We were taking the concept of a DLP solution, which was edge-based technology, like your firewall, and moving it down to the endpoint itself.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: What should an organization prioritize first to stop AI data exfiltration?

Williams describes three phases. First, audit to learn the size of the shadow AI problem and which tools are in use. Second, decide on enterprise licensing, legal agreements and a basic policy. Third, enforce with anti data exfiltration technology that measures prompts and governs with real controls, because people will revert to the tools they like.

“You ought to audit what’s actually happening before you do anything.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: How can you protect sensitive data without slowing employees down?

Williams says ADX Vision 2.0 was designed to run hidden in the background and intercept prompts going out to any frontier LLM. If a prompt contains something like Social Security numbers or credit cards, it blocks the prompt and asks the user to rewrite it. Teams can add custom prompt management for their own intellectual property or language-specific content.

“We are not trying to stop you using AI. We are encouraging you to use AI.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: How do you detect stolen API keys and token theft?

Williams says stealing other corporations’ API keys to get free AI is a growing trend, since AI can be expensive. BlackFog’s tokenization determines when API keys may be stolen, and a unified dashboard shows when they are used out of bounds or by people who should not use them. Teams can add custom detectors to keep internal information, such as databases or IP addresses, from being pasted into LLMs.

“One of the greatest trends right now is stealing other people’s corporation API keys and utilizing them yourself and getting free AI.” Dr. Darren Williams, Founder & CEO, BlackFog

Q: How does the EU AI Act affect AI governance?

Williams says the EU AI Act is now in play, and it requires audit capabilities and certain checks and balances. He expects similar rules to reach the United States at some point.

“There’ll be variations of that, such as GDPR, but specifically applied to AI itself.” Dr. Darren Williams, Founder & CEO, BlackFog

Resources & Documentation

  • ADX Vision 2.0, a BlackFog product that monitors prompts sent to AI tools
  • ADX Protect, a BlackFog product in the ADX family

***

👇 Click to Read Full Raw Transcript

Swapnil Bhartiya: As the adoption of AI is growing, we all know that people love their tools, people love their platforms. So employees are pasting customer records and study documents into AI tools that nobody in their organization actually approved. Now that data leaves through a text box, one prompt at a time, past every control that was built for files and network. And now AI agents are acting on corporate data on their own. So there may be no human in the loop To catch a leak. And to unpack shadow AI, agentic AI, and how to stop data leaving the endpoint, I am joined by Dr. Darren Williams, founder and CEO of BlackFog. Darren, it’s great to have you on the show.

Dr. Darren Williams: Great to be here. Thank you very much, Swapnil.

Swapnil Bhartiya: It’s my pleasure. Before we get into this problem area, as we were talking before the interview started, the company was founded around 2015. It’s 10 years then, but time flies by so fast. What I want to understand from you is not that why you created the company. Yes, that is a good story as well. But when you created the company, ChatGPT and all those things were not there. I mean, AI has been around for ages, but these LLMs, they are new phenomena. So how much things have changed from 2015 to 2026 and how the company’s focus changed and how you also felt that The reason you built the company is actually more relevant today than it was back then.

Dr. Darren Williams: Very accurately put, Swapnil. Yeah, I think it’s really important to understand the foundation of the company was really framed around the idea of data movement and data loss. You’ve got to step back to 2015 and where we were. This is even pre-ransomware, if you think about it. We were actually concerned at the time about data moving off the device. There are a lot of tools around that really focus on attack vectors, which are endpoint detection and response tools. That really focus on, you know, stopping the bad guy getting in. Black Fog’s always about stopping the bad guy getting out with your data. Now, this was interestingly before, as I said, before ransomware even started. And then back back in about 2019, we just started to see this trend of you know that used to be that ransom was really focused on encryption. They moved really quickly around 2021 into. Basically not encryption but really data extraction. And really that was perfect for us because, you know, the first— you got to remember the first 4 years of the company was really about perfecting a technology. We were sort of taking the concept of a DLP solution, which was edge-based technology, basically like your firewall, and moving it down to the endpoint itself. So our technology really sits on the endpoint, which is where all that— all the, uh, all the data, you know, extraction happens. And basically we had a technology which was miniaturizing without breaking the secure context of the protocols themselves. So DLP, traditionally it’s focused on basically unencrypting your SSL traffic and be acting like a man-in-the-middle attack. And so we thought we can do it with other technologies. Now we have the techniques available to stop the bad guys removing your data. So that’s the genesis of the concept of the company. And that was really interesting. We’ve got a lot of customers. Our base product is really about, you know, stopping attackers and ransomware effectively. And then of course we saw the evolution. And by the way, we call that anti data exfiltration in terms of a technology. Then you saw the evolution of LLMs, right? So it wasn’t really until about 2024 where we started to see LLMs and OpenAI really take the mantle. And then it just exploded, as we all know right now. And interestingly, it’s still a similar problem, right? The difference is now that data is leaving the building through the use of LLM prompts. So prompt management became a really big thing. And if we fast forward even further, we currently solve the problem of LLM data loss because we’re basically— it’s really an uncontrolled environment. And shadow AI, as you mentioned in your introduction, is a big problem. Is probably one of the biggest threats to most companies out there right now. Because if you think about what companies are doing, they’re trying to adapt to this new world that we live in, and they’re still testing. They don’t even know how they’re going to be doing it. Some are using it content development, etc. But with that, we have a lack of control around what’s happening. It’s developing so quick. We’re losing control of the data and we’re losing control of corporate intellectual property, if you will, as well. And so our technology was extended beyond our genesis to focus on LLM prompts as well. So that’s really what we’re focused on. But, you know, I’ll leave it at that for the moment so you can ask some other questions. But I just wanted to show you the evolution of where we are. It’s just that the world has changed But our technology was adapted to really fit that new realm that we’re in today.

Swapnil Bhartiya: How serious is this problem as employees bring in AI tools that in most cases their employers never approved? Even if they said, hey, don’t use ChatGPT, they might end up using it because easy or they’re familiar with that. What are the biggest securities that it creates for the organizations which you were like, In pre-LLM era, those risks were there, but not that bad.

Dr. Darren Williams: Yeah, no, it’s a very good point. And yeah, this is the major deal that we’re focusing on these days is, you know, shadow AI is really about discovering what is actually going on in the walls of a company. And most companies don’t actually know. You can mandate all you like that, oh, we are a Copilot shop, we are a Claude shop. But the fact is people will revert to the tools that they prefer And use themselves, and so you can put in great policies. But if you’ve got no controls over it, then you really don’t really know what’s going on. In fact, very early on, like about a year ago when we released our first tool in this space, we did an audit for some companies, and they said initially, “Oh, we know exactly what’s going on.” Fact is, they didn’t at all. They did not. They were devastated. They did not know that. We, but we have a policy in place, so surely people follow it. No, they don’t. They open ChatGPT, they open, they open whatever that they’d like, but they’re even more shocked when they see the data that is being pasted into these non-approved LLMs. Now, many of them free. Now, it’s really interesting to me, the gut— you, you mentioned governance as well, we should touch on that. The problem you’ve got with these LLMs, not only is the additional skill sets and the connectors that you mentioned, but the fact is that they all have different policies around what that company will do with your data. So if, for example, you are pasting some, you know, confidential, uh, IP into ChatGPT free edition, they are free to use that for training purposes for their LLMs in the future. Now, once you’ve done that, you’ve lost that data forever because of course it’s now assimilated into the Borg, if you will. And you’ve lost that intellectual property because now I can go query and it will come back with the right answers and it will say, oh, give me all your intellectual property from TFiR and it will give it to me, right? So these are the problems that we’re focusing on and it’s very nuanced, right? So the legal contracts that you make with these LLM vendors vary significantly based on the tiering of that. And on it, in addition to that, The settings, which I don’t know if you’ve ever looked at the latest settings on OpenAI or Claude, they’re extremely complicated. There’s literally half, you know, 20, 30 different options to control your data security. That becomes really difficult for people to even know what to do to protect their data. So let’s frame that as the shadow AI problem, right? We don’t know what employees are using. We need to take stock and audit. The next level I think we need to touch on is agentic AI. You mentioned the connectors before in your comment. I think this is the new frontier. You know, the technologies are developing so quickly, people cannot keep up, and the security definitely doesn’t keep up. And so what we’re trying to do here is we’re trying to work out exactly what we can do to stop agentic AI because For us, we have agents that are running and doing things all day long, right? We’re doing it from from a security purpose, but other companies are doing it and saying, “Oh, make a phone call, book this restaurant, communicate with customers,” etc. And you don’t really know what these agents are asking of the LLMs, and so we have no visibility. So not only do we have a shadow AI problem, we have an agentic AI problem, which I don’t even think people realize how serious this is going to be because. It’s going around making decisions on your behalf, sending stuff. It’s got access to, say, your Google Drive— classic example. What’s it doing with Google Drive? I don’t even know, right? Because you don’t have visibility into that. So we need some governance around it, or controls and constraints that says you can’t go and download all my contents of my Google Drive to help train the engine. And so that’s what we’re trying to stop here. Yeah. And, you know, it’s a cat and mouse game, like you said, quite, quite true. And we’re just trying to stay on top. This is what we, we wake up every morning. This is what we focus on. So it’s a really fascinating area, but it’s also like, you know, it consumes all of our time.

Swapnil Bhartiya: Thank you for that. Now, I also want to talk a bit about how serious is this problem of, you know, unapproved AI tools, because you folks, your research with Sapio found that almost 49%, I think, of employees surveyed have used unapproved AI tools at work. What does that number reveal about shadow AI and the exposure of sensitive data and in a way inability of organizations to actually force those policies? But why, why is it happening?

Dr. Darren Williams: Well, people revert to the tools that they prefer. Some people use Bing, some people use Google, right? And people are very attached to their tools of choice. We see it every day. We see it in You know, every type of environment within a security organization. It’s people you can’t move them off. It’s very sticky, right? And oh, I know how to use ChatGPT. That’s what I’m going to use. I don’t care if my company says we have to use Copilot. They would rather take the risk and get their work done faster because that’s what they know than actually use what the company mandated policies are. Now, problem is we don’t have the controls around that yet, right? That’s why we developed our product, is to give the company visibility visibility into what’s going on and actually allow them to automatically redirect. So if you use, say, OpenAI, but we’re a Copilot shop, it automatically redirects the request to Copilot so you can’t get around it. And that’s really what that governance level is all about. And then providing visibility into the prompts themselves and what these, you know, what these tools are doing, because, you know, we’re Data poisoning is a great example, right? Data poisoning is a new concept where you basically can go in and extract content from your hard drive and send it up into the LLMs, and you can encrypt it. You can, you can capture, obfuscate the data as well. We’re trying to expose what’s going on there. I mean, we’ve seen all the attacks going on, but my big concern, frankly, is the The fact that these engines are getting so intelligent now, they’re able to break out of their four walls, right? They’re out of their environments. We saw that with Hugging Face. We saw that with all sorts of attacks that are going on right now. They’re so smart; they’re able to bypass the four walls, and that brings up a whole new range of problems and concerns for me. Just a member of society, let alone in technology. Like, what is this going to mean for the world? I mean, it’s a great power. Powerful tool, but at what cost?

Swapnil Bhartiya: So very well said. Thank you so much. And I want to just continue the discussion on autonomous AI agents because they can access, as you mentioned, process and share corporate data without a human watching every step. What new security challenges does that introduce?

Dr. Darren Williams: Well, that’s sort of what I was referring to. The agents and the, you know, the autonomous nature of them is getting Way beyond what we even thought it was going to be, and I would say you’re probably only dealing with you know five to ten percent of the world is really actually experimenting and playing with this right now. So the numbers are still low, but this is the next frontier from an AI security perspective because you know people don’t know what it’s doing. They want to automate tasks. They want to get work done really quickly. They’d rather go watch the kids play soccer than sit there writing these big essays and content for the organization. So they will use those tools and then they will go and ask them to go and research topics and correlate data from disparate systems. I mean, it’s having a huge effect on a lot of, you know, business intelligence companies as well because this is their bread and butter. This is what they did. Now we don’t need them at all because these autonomous tools go out and do the work for them. But again, what— at what cost? Here as well, because if they’re going out and getting and, and getting that data because you’ve asked it, what questions is it asking? What are the API calls that they’re making, and what intellectual property is being lost in that process? We don’t rightly know, and we’re still adapting to this new environment. And these are just today’s problems. Tomorrow’s problems are going to be even greater as they even get more intelligent. And there’s more connections. It’s, it’s a fascinating area, great to be in, uh, the environment. But like you say, it’s a cat and mouse game and we’ve got to try and keep up.

Swapnil Bhartiya: Let’s also talk about— you touched upon that in the beginning, you talked about DLP and all those things. Why are traditional security controls like, of course you mentioned DLP, EDR, uh, they are— why are they struggling to keep pace with generative AI and agentic AI?

Dr. Darren Williams: Well, I think the problem is they were developed for a different place in the world, right? We weren’t in that space. I mean, you think about before that, even if you think About the antivirus market, you know when there was a few thousand viruses, no problem. Take a fingerprint, do a scan, and Bob’s your uncle, right? We’ve fixed the problem. But it developed so quickly that that technique that they used, this classic fingerprinting technology, was not able to cope with the sheer volume. And and then the attacker adapted, and the attacker says, “Well, okay, if you’re going to do that, we’re going to put on gloves and effectively have no fingerprint.” And so polymorph. Code developed where the fingerprint would change every time it gets installed. So that technology became pointless, right? So where we are today— so EDR developed from that because fingerprinting doesn’t work anymore, right? So they developed a new technique. And again, we’re now at another technique and another big change in the world because AI wasn’t even around when EDR was invented. And so the techniques that they’re employing to do this don’t work. They’re not even focused on— because AI is all about data. And if there’s one thing I’d like your audience to get out of this is that it’s always been about the data. It always will be about the data. Even the attackers know that. That’s why they change from encryption-based technology and bricking your computer to stealing your data and extorting you with that data. And AI changes it again Except that we’re getting the data using AI instead of traditional techniques. So these traditional tools were not built for this new world that we’re in. We have to fight AI with AI. It’s really interesting to me when I, you know, I’m constantly reading and scanning the news every day. I think it was from The Economist this weekend. There was a really interesting graph that they put out. You’ve probably seen it already. Which actually tracks the amount of AI use in AI companies. And if you start— if you go back for 3 or 4 years, the amount of human involvement in training AI was significant, like 80-90%. If you fast forward to where we are in August of 2026, 98% of All training is done by AI engines themselves, so it’s adapted to their own self-learning. So it’s the human involvement in this to train it and to get better is so minor now comparatively to where we were, because the engines are now so smart they can train themselves. So I got great stories of some of that too, if you’re interested, just in terms of the way these companies are actually utilizing software engineers. But maybe that’s a subject for something later. But it’s really interesting The way they’ve adapted. So even the AI companies themselves have changed the way they work. So back to your original question, yeah, the original tools that were developed were not developed for this world that we live in. That’s why we developed the, you know, the focus on AI from the ground up.

Swapnil Bhartiya: Let’s talk about what if an organization is— first of all, no one is starting from scratch, but the pace at which AI is progressing, as you rightly mentioned, a lot of organizations are, you can say that they’re reinventing the wheel every other day, how to deal with AI. So yes, we can use this phrase that if an organization is actually starting from scratch on this problem, what should they prioritize first to prevent unauthorized data sharing and data exfiltration? Let’s talk about the practices first. And of course, we’ll talk about BlackFog as well.

Dr. Darren Williams: Sure. I would say you’ve got to start with We talk about these things in 3 phases. First, you’ve got to know the extent of your problem, right? Just when you, you basically ought to audit what’s actually happening before you do anything, right? What is the scope of the problem? Scope it out just like a consultant coming to your business. Let’s have a look at what the problem is. So you’ve got to audit that and you need a tool to actually tell you what is this, what is the size of this shadow AI problem in my organization. Then find out what tools are being utilized by your organization. Are we use? Are we 80% Claude? Are we 20% OpenAI? What is it? Once we’ve got a foundation and we’ve got the data that we need, we need to now work out what are we going to do about it, right? Are we going to license these versions at the enterprise level? What are the legal agreements in place? And set out a basic policy for your organization. If anyone’s interested, we help you guide you through that process on our website. Website, and we’ve got a few articles about that. But basically, we really think about how do we govern this, because you don’t— you want your company to be effective, you want it to be fast, you want it to be adapting to this new world that we live in. So you don’t want to slow them down, but you do need to have the adequate policies in place. This is particularly important in Europe with the, the EU AI Act, which has come into play now too, where you have to actually have audit capabilities, you have to do certain There’s certain checks and balances that they’ve actually added in place. I expect this will also start coming into the United States at some point as well. So there’ll be variations of that, sort of like GDPR, but it’s with AI itself. So what— that’s stage 2, right? So have a policy, have some level of governance, and then thirdly, you have to actually manage it with a tool, and you have to actually have some anti-data exfiltration type technologies in there. So you’re measuring the prompts and you’re governing it with real controls, right? It’s great to have a policy, but if you’ve got no teeth and you can’t actually control it, then you— what have you really got? Because as we pointed out earlier, people will just revert to the tools that they like to use. So you have to be able to have sort of checks and balances in place to actually stop that happening and to make sure they’re adhering to the policy, frankly. So really, that’s what it’s all about. So that’s how we think of the 3 phases of growth within the company. And this is becoming really important— that 3rd phase of, of the governance and enforcement layer is becoming really, really important with agentic AI and the rise of that, which I expect in the next 12 months will be the biggest thing that probably has to be addressed in the world really in terms of AI security.

Swapnil Bhartiya: Let’s talk about your solutions. Of course, the whole ADX family, Vision 2.0, and so many things are there. Many CIOs worry that locking down AI use will kill productivity because it is actually helping a lot. So, I mean, as we have seen, we have to maintain the balance, give it, I mean, we do talk about guardrails, we do talk about, you know, the whole human loop, AI is smart and it finds its way. How does ADX Vision protect sensitive data without getting in employees’ way, without taking away their flexibility and freedom to be able to use AI?

Dr. Darren Williams: Yeah, so from the onset when we designed this, we basically focused on a tool which is basically hidden but in the background, and you don’t have to do anything. You basically deploy it and it just watches in the background. It’s like a monitoring solution, and it intercepts all prompts going out to any LLM in the world, right? Any of the frontier ones, at least at this stage. And so the idea is once you type something in OpenAI, if it detects a problem with some of the prompts and there’s, there’s some issue with it, it will block that and tell you, hey, this has some, um, secure data that we don’t want you delivering, uh, please rewrite your prompt and stop uploading confidential information. So it scans for Social Security numbers, credit cards, whatever you want it to do, and then you can actually add your own custom, um, prompt management in there as well based on your intellectual property or your language-specific stuff in your organization. So we really focus on a very seamless environment. We propose that we are not trying to stop you using AI. We are encouraging you to use AI. What we’re proposed— what we do is we put the guardrails around there so you as a corporation Do not leak data that you don’t want leaked, effectively, because we understand we use it every day, right? I mean, we’re some of the biggest users of AI out there, right? Because we’re using it for security as well. But we don’t want to be in the way. We’re not trying— it should be seamless, it should be touchless, it should just be the guardrail that you need it to be. And that’s what we’re all about.

Swapnil Bhartiya: In addition to AI, you know, breaking all the boundaries and do all the bad stuff that it can. There are a lot of other things which are, which has nothing to do with AI, but it could be like, for example, APIs, API keys are leaked, credentials are leaked, or let’s say token theft is also there. What are the other emerging attack vectors that you are seeing? How does it look like in practice and how can BlackFog help with that? Mitigating those risks as well.

Dr. Darren Williams: Yeah, so one of the parts of the product that we, um, we spend a bit of time on is actually the tokenization part of the product where we’re able to determine when API keys may be stolen. That’s a common one, right? Because AIs can be pretty expensive, as we all know, right? And you multiply that out, one of the greatest, uh, trends right now is stealing other people’s corporation API keys and utilizing them yourself and getting basically free AI. And so that’s becoming a bigger problem out there. And so we’re really focusing on protecting those API keys and providing a unified dashboard where you can see when it’s out of bounds and when it’s actually being used by individuals that maybe shouldn’t be using it. And that’s— there are a few examples of that out there in the workforce. So yeah, that’s a really big problem, scanning API keys, any API keys for other things as well. So you can actually add your own custom detectors in there as well, where you could detect various things, maybe some internal corporate information where you could actually protect it from accessing certain databases, IP addresses, etc. So we want that to make sure that does not get pasted into LLMs. So again, another important aspect of the, of the technology.

Swapnil Bhartiya: Darren, thank you so much for taking the time to walk us through this, to talk about, of course, shadow AI, the whole problem area, and how BlackFog can help organizations mitigate some of those challenges. Thanks for your time. And those who are watching, please go and check more about BlackFog and ADX Vision 2.0 at blackfog.com. And back to Darren, thank you for joining me today, and I would love to have you back on the show. Thank you.

Dr. Darren Williams: Thank you very much, Swapnil. Pleasure to be here.

How to Stop Guessing Kubernetes Resource Requests | Shahar Azulay, groundcover | TFiR

Previous article