The Reason 30 Years of Cybersecurity Has Failed - and What Actually Fixes It | Trent Telford, Qanapi

0:00As you point out, once you're inside the wall, you can see everything. People started building or buying security systems that were effectively a wall around firewall with doors that would go into it. And they kept the doors locked unless you have the proper permission. It doesn't matter how high you build the wall. The bad guys go get a bigger ladder. Well, they find out how to pull a few bricks out of the wall in the bottom of the castle wall and crawl through. And you don't notice that they've crawled through, right? They pull a rock back over and you don't know they're in there. So that's been a really big issue.

0:30We'll get to AI, but particularly with these mythos classes that can scan a code base or a piece of software externally and find undiscovered holes that then they can write exploits for us. What's the solution that you guys have come up with? We start from a baseline assumption that the data will be exposed. Let's start by having you introduce yourself to listeners. and explain how you got to Quantipi. And for listeners, it's Quantipi as in quantum API. Is that right? That's right. So you can explain that. But give us your background. I know you had a startup before this. Yeah, thanks, Craig. Thanks for having me on.

1:18And it's great to be here. Probably 20-odd years now, 25 years in the tech game, particularly in software and internet related. So I come from that background of 25 years originally in the late 90s, working for the big investment banks, but on big transformation around e-commerce and technology projects, building internal intranets, as we called them, as we all remember in those days, for trading systems and things around Bankers Trust. That was where I started. And then when they fell over, I ended up with Deutsche Bank. And then in the 2000s, I was working in e-commerce consulting. So it was a great place to learn all about, you know, the full stack of what happens across as you want to design the internet-based systems, which we all take for granted today.

2:11But, of course, 25 years ago, that wasn't the case. And then I almost fell into the data security side probably 20 years ago now, nearly 20 years ago, 19 years ago. I was doing some implementations of big identity and access control systems across banks and government, and then came into the very specific area around data security. And that happened back then. So I've sort of been looking at this problem around data security and all these issues for so many years now that what now people are looking at and thinking, wow, that's a great idea. That's logical, right? That's logical. That's simple.

2:51is something that simplicity has taken us a long time to get there over various iterations and various startups over time. So, yeah, and I'm now, despite the Australian accent, here in the U.S., and I live in Washington, D.C. Okay, let me give a little recap as I understand the history of cybersecurity or protecting enterprise systems. You know, initially, for a lot of people, it was just making sure your databases, all the ports were closed and the proper permissions were in place. I know that S3 buckets, which is a popular data storage thing, you know, there's, I had a hacker that was showing me, you can scan the world for S3 buckets that are misconfigured and pop right in and see all their data.

3:52You know, it was amazing to me. And then people started building or buying security systems that were effectively a wall around a firewall with doors that would go into it. And they kept the doors locked unless you had the proper permission. But as you point out, once you're inside the wall, you can see everything. So then I've seen, I mentioned to you a few years ago, I was talking to a company that was developing data layer security, but they were kind of a guardian that sat at the door to any particular data store and would verify anything or anyone going in or out. but they didn't protect the data itself that was going in and out.

4:56It, again, was a permissioning thing. So you guys are taking a completely different strategy or talk. Can you tell us what that is and ultimately how it works and why, I mean, what is the quantum in the right quantity? in the API. Yeah, look, I think if we take a step back, I mean, now you're hearing a lot about zero trust at the data level, right? We've seen zero trust as a general, what started as a trend that then became best practice in applications. So if we take a step back, like you say, I think what people need to remember is the internet was never designed to be secure. It's actually designed to share information from when, go back to the DARPA days of invention and then the initial, you know, putting up a web page for information, you know.

5:49And then people, you know, just for a bit of a quick history lesson, I mean, SSL, you know, the encryption of the connection from your browser to a server that we all take for granted now with the little padlock. Mark Andreessen, the famous, obviously, American venture capitalist, was one of the first and correct, I'm sure I'm quite correct in saying he was the one that invented that back in the early days, right? And, you know, we still use that today. It's still a network-based security. So if you think about it, it was never designed to be secure. So therefore, what we've spent, let's call it 25 years, 30 years doing now is kind of retrofitting, constantly chasing our tail to get ahead of the bad guys, to try and put more walls up, put more moats up, put more protection, better locks on the doors, all that sort of analogy.

6:42the problem is it doesn't matter how high you build the wall the bad guys go get a bigger ladder or they find out how to pull a few bricks out of the wall in the bottom of the castle wall and crawl through and you don't notice that they've crawled through right they pull a rock back over and you don't know they're in there so that's been a really big issue and as you say something needs to change because if you think about where we came from with traditionally PCs and on prime and big mainframes. We then moved to internet. And then in, you know, 07 along comes Steve Jobs and, you know, he has a smartphone.

7:18And now we'll talk about this a little bit more, but all the AI side of things and all that autonomous and drones and landscapes, everything else out there. I mean, anybody, whether you're a tech person or not, you get the picture. Like you can visualize all those drones in the air, all this AI, everything. And these are borderless. They are literally borderless. There is no boundary. So the old, you know, put there to be a lock on the door, that model's in real trouble. Yeah, and also with regard to being borderless, you've got multinationals. I mean, there's no longer a locality that you build the wall around.

7:55It's dispersed.

8:00We'll get to AI, but particularly with these mythos classes, what some people are calling them, these new models, Fable and Obus V and Mythos V, that can scan a code base or a piece of software externally and find undiscovered holes that then they can write exploits for. So that castle wall and the guy pulling the brick out of the base and crawling through, that's becoming a real problem because not all the bricks are secure or they're discovering that. So tell me, what's the solution that you guys have come up with? So we looked at it differently, and we start from a baseline assumption that the data will be exposed.

8:58You know, I mean, I think there's nobody who works in cybersecurity who, well, certainly nobody is credible in cybersecurity that wouldn't say that it's not a matter of if it's when you get hacked or there's a data leak, either intentional or unintentional. And by the way, more often than not, it's unintentional, right? But if you look at the scenarios, we assume that there's going to be an Edward Snowden event, right? We assume that, you know, remember Edward Snowden was a very trusted person. He had all the credentials and he just decided this needed to be out in the wild. So whether you come in and attack through a vulnerability found in the wall or through trusted sources, we just assume that the data will be exposed at some stage.

9:40So the first thing we did, we looked at encryption, and most encryption is not done at the data level. Yes, you can encrypt a file, but mostly people don't do it because it's useless once it's encrypted, right? and it tends to be if people do it at the data level, which is rare, it's a wrapper. So typically all content and data sits in databases open because otherwise the databases can't read it, they can't do their query job. Or unstructured data sits on big servers like S3, if you like, AWS as you rightly port out storage. So we looked at it differently and thought, well, what if I can encrypt individual parts of a file?

10:21So if you've got a Word document or you've got large flat files or unstructured data, we can get down to individual words, paragraphs. In databases, we can do individual cells, and we apply a unique key for every word or paragraph. So we're right down at that level. So we assume straight away that if you've got sensitive data, you can almost portion it into all those encrypted fields or data sets. But the big problem with encryption, so first of all, people hadn't thought of it at that level. The second issue is that people hadn't thought about key management tied to identity and policy. So I try and explain that and say, well, when you walk out your front door in the morning, Craig, do you get 10 keys and just scatter them all over the footpath or hide them under the mat?

11:07Right? Under the mat or pop and hope that, you know, that the bad person doesn't find those keys. No, you give one to your cleaner or your, you know, wife or your best mate and say, look, come past the house. And when I put a lot of that, people say, oh, that's logical. Well, that's what we did. So we said, let's tie identity to that key. So identity can be a human, obviously, and you present your credentials as you do today, two-factor or whatever it is. Or machine, a drone. We know what a drone is. It has its hardware credentials on there or a machine or a server. They all have hardware credentials.

11:42So if we tie the key to that identity, then we know that only the proper person or machine can have access to that key for that data set. And then we realised the third thing was, well, we could set policies around it. So it might be you prove your Craig. You prove that the system says you are allowed to have that key for that data set. But we're going to add some conditions. It could be time, only between 10am and midday. In the case of drones and autonomous landscapes, it could be temporal. For example, GPS plus altitude plus drone identity equals yes. If all those things come together for an encrypted payload on the drone, encrypted payload 56, you can get key number 56, simplicity.

12:26So tying those three things together is unique, and nobody else has done it. And that's what allows us to build all sorts of solutions and deliver all sorts of interesting things that nobody else has done. Yeah. On key management, I mean, these are not keys that people are storing in their Apple wallet or something. These are keys in the system. How does the person, if I have a key tied to my identity and I'm working on a system, how do I give the key to the system or does it have the key stored for me and my identity is validated? Right. So, yeah, that's a very good question. So that's coming back to the name, Quantum API.

13:18So it's an API service. So our customers are large enterprises, US federal type markets, defense markets, where they integrate it into either existing platforms. So they integrate it with the software that runs the drones, for example, or they integrate it with the software that manages all the files, for example. So when you set it up, you might have, you know, you want to protect certain data sets. You can start by adding, you know, you might point our service to Ping or Okta, who are your identity credential multi-factor providers for your organization. So now when that data is encrypted across your organization at, you know, individual levels, you don't actually, it's kind of transparent is the word to you.

14:02You just know, for example, in Google Docs, I'll give you an example. In Google Docs, we can do multi-level classification in Google Docs. So we already know who you are when you're logged into Google Docs in your enterprise, right? We know it's great. Now, we might know that you have, you know, top secret clearance, right? But your pal Billy only has sensitive clearance. So when you're in our docs, you see individually encrypted different paragraphs. It will give you, it will automatically, when you go over it, decrypt that part that you can read. But Billy goes in, he can't see that part. He can see the next layer down, but he can't see the other one.

14:40So it's integrated as part of the processes as an API service. And that's the secret sauce. It's not the individual running around trying to manage the keys. It's enterprise grade. Yeah. And when we were talking earlier, it sounded like, so you have a document that's in a data store of some sort. And when you call up that document, it automatically encrypts down to the cell level or personal identifiable information or different strings in the document. That happens on the fly, or is it encrypted in the data store? It can be either or. So it can happen on the fly as part of a gateway service at massive volume.

15:37So it depends. You can do it on the fly through a gateway service. Or you might be looking to do an uplift because you're a federal government agency and there's mandated timing for post-quantum now between now and 2030. So you might go in and say, I've got a million files sitting in there where there's sensitive information in these files. because it's an API, we can write a bit of code, an SDK code, keep it simple. It might be Java or Python or whatever. And it will go through that and you can run it almost like a batch, if you like. And you can go back and over time, depending on how long you've gotten, the compute power, it'll go through those million files and automatically encrypt all those individual cells or fields or words.

16:17In the data store. So in S3, for example. And then obviously that data at rest is then protected. So if you come in through the ways we talked about, so either if the insider tries to take it, well, good luck. What are you going to do? You're going to walk out with a million files with millions of little encrypted words. Bad guy gets in. Same problem. Same scenario, sorry. And equally, it covers that data at rest, but it also covers the scenario of data in transit because now it doesn't matter. You can send that file over open networks, right? Untrusted networks. because assuming you've protected the sections you want to protect, the clear text you've assumed is public anyway, so everything else is encrypted.

16:58So now data in transit is covered and data in use is covered because, yes, it was stored, it was transited, but then it gets used. So let's say the Google Doc example, because you've got 100 people looking at it with different clearances or in an enterprise that might be board level or C-suite or whatever, research versus sales in investment banking, in use means literally while it's being used, you can only see the right bit. So it covers all three states of data too, which is highly unusual. Yeah. And that decryption, it happens. Where does that happen? So I'm at a workstation. I call up a document out of the data store.

17:43Or as you said, the Google Doc, when I see it, my credentials allow me to see certain things. Someone else would not see certain things. Does that happen on my computer? Yes. Yeah. It happens in the browser. So if we continue on the thread of Google Docs, it actually brings up sort of a pain on the right-hand side of the screen. So actually what you see is encrypted text when you sort of click or roll over it. On the right-hand side in a little window, you'll see that same paragraph, for example, in plain text format. And the reason we do it like that is that means if you – because if you decrypt it, A, it turns up in the document history, which you can clear, but it's in the history.

18:25But secondly, it means it's touched Google service. Imagine that that's the Google Doc connected to Google that way, and our add-on service is now connected to our service. So it never lives in clear form on that Google Doc. It happens client-side. So when you encrypt, it's creating brand new keys. It's sending them down to the browser. It's 100 keys for 100 different paragraphs, for example. It's automatically encrypting all those client sides. So now it's fully protected. But there's no footprint. That's an important point. There's nothing to install. There's nothing to run, which means it runs across mobile devices.

19:01It runs across anything. And that's the beauty of being an API service. Yeah, yeah. And I was asking just before this, what about false positives? I mean, what's the accuracy in reading the keys or in encrypting specific data sets? You know, look, it depends on the data sets. The short answer is we don't make it, we, our tool doesn't make a judgment on the efficacy, if you like, of that. Obviously, if you've got structured data, it's much easier. If you've got a big database, you know what the different fields are. As long as you, if you're using AI, you're using our tool on the client side in AI to go and look at the training of that data to make sure it protects all the sensitive fields.

19:52If it's in structured format, it's much easier. If it's unstructured and there's millions of files, but every file is the same. So let's say it's an insurance form. Forms are a good example in PDF, full of PI, SSN numbers, all that sort of stuff. That's pretty easy, right? Because you can train it, make sure if every document's the same, now I want to push all that into a big frontier public LLM, I'm going to have high efficacy. If you've got a million files that are all slightly different, obviously the chances of missing something gets bigger, but that's really just an exercise in training in the AI example.

20:26Yeah. And that training, so when an enterprise calls the API, your model is already trained on every configuration of sensitive data that could exist, right? So we have a small footprint, effectively an AI agent that can sit out on the client side. So let's call it on-prem or in a client's environment. And we don't have very specific ones in terms of knowing what's sensitive is not. You train it. So you go through and you put the document in and you train it to make sure that it captures everything. You look at the before and after. You make sure it's all happy. And effectively, you save that scenario.

21:13and then it can then automatically go through and any data that's being pushed up towards LLMs is automatically encrypted. So, you know, you and I talked about this previously at another time, but that's a big issue right now with all the public LLMs, right? The big argument is we're going to have to build this stuff, you know, on-prem. We're going to have to go and buy our old infrastructure like the old days and suddenly have racks all through our offices of servers. And, of course, the biggest topic in the stock market, right now is have we over-invested in AI data centers and AI capacity? Well, this is one of the big topics, right?

21:50If you can find a way to have enterprises move vast amounts of data up to public LLMs, then we can talk about the other usages of that and what other drivers will be. But that certainly helps a lot, right? So nobody wants to go and buy their own hardware or have to go and rent their own. because even if you go build your own or rent your own private cloud infrastructure, you know, I can go to Google Cloud and, you know, rent that AI infrastructure, you can run your private models there, but that's going to be a compromise, right? Because your context models are much smaller, your context windows are smaller, your grounding data, all sorts of things mean you've got a very different output to what you do in the big public LLMs.

22:33And so that's one of the big issues we think we've solved is allowing companies to, we call that negative provenance. So you can encrypt all the individual bits of data. It actually comes from the organisations on-prem up through our gateway. So we check all that and we don't let it go into your AI service, whether that's Anthropic or all the main ones, unless all that is ticked off. We're a gateway service. But the second thing we do is the keys to encrypt the originating data, we know that identity, right? We know it's your data lake or this human's machine. So we call that positive provenance.

23:10That means it's all hashed. It's encrypted with an identity when it was created. So we know when it goes to the gateway that it came from your trusted source. That's a massive issue in AI right now. What's a trusted source and what's not? What's poison data? What's not poison data? So solving the positive provenance, it is my data, and making sure it can't be tampered with. And then secondly, solving the negative provenance, making sure sensitive or classified data didn't get into big commercially sensitive data. Public LLMs is a fundamental issue. But because we've got that foundational identity, key, and policy, we know when we send the keys down, the request came from a trusted identity, Data Lake or the trusted S3.

23:52We sent the keys down, all the policies meet, and away we go. Yeah. We were talking, or I mentioned to you, there was a story the other day about chat logs from public LLMs showing up in search results, which is pretty shocking. But in this case, if that happened, and I think early on, right at the beginning of chat GPT-3, Samsung, there was an incident where some of their private data was uploaded to an LLM and then exposed. So this would, if anything was exposed, the sensitive cells or strings in that data would be encrypted. So no one would. So it's useless. Yeah. I mean, and that's the thing.

24:48You've got to make a business decision what's sensitive and what's not sensitive. We will encrypt it. But, you know, you can literally see it goes through the gateway. It goes into, you know, we've got a demo example for people to play with. It's Claude, for example. And Claude literally says, I found this, this, this, and this. I couldn't read the data in this section because it appears to be encrypted with an encryption key. And then it goes on to give you the results from the clear data. So it literally can't read it. So there is nothing to be concerned about then. That literally is, that problem is not there any longer.

25:21It's that simple. And in terms of the model's performance of the reasoning, or whatever you're doing with it, it can still do that. It's got all the language around the encrypted strings or encrypted cells. Right, right. Yeah, and I mean, of course, that's always going to be a trade-off. If you look at what you're wanting the AI to achieve for you in terms of insights, you know, if the data you're encrypted has nothing to do with those insights, then obviously you'll still get the same, essentially, efficacy of outcome. If you have to protect certain fields, So, for example, if you're a large insurance company or bank and you wanted to use public LLMs but you're not now, or retailers with loyalty programs, the social security number of Billy or Mary is going to have absolutely no bearing, obviously, on the outcome of what you're doing.

26:15So no problem encrypted. If you were to protect lots of other information that may have a bearing, well, you're going to have an adjustment in that weighting. But, you know, I think we've got a customer that's a very, very large retailer. And they have a competitor that's a very, very large retailer. And I know one of them is using it very strongly, AI, and the other one is simply not using it because of these issues. And I think, you know, the question for the big enterprises is how long can they last abstaining or staying off the train, right, before they get blitzed? What's the trade-off between competitive advantage or death quite quickly by not moving with those insights?

26:55And I think there's a lot of companies struggling, rightly so, with that issue right now, right? Yeah. This is an API service, and this is for enterprises that are using foundation models through an API, right? It's not desktop. Right. It's not for me sitting or someone sitting just banging it into chat GPT. We are looking at that, but that's where it goes. You're typing straight into your chat window. You're going straight into CORD, for example, on your desktop. We built a gateway service. There's big enterprises or organisations and the data passes through us. You give us, when you sign up, you put your API key for, say, you use four different AI services, and that data goes in.

27:47And you get a visual representation of that chain, if you like, of everything being Providence chain encrypted. Yeah. It could work by even on a desktop or working with one of the model apps. you could have the system check the prompt before it hits the model. Yeah, we haven't looked at just yet some of those. But yes, in theory, there's no problem with doing that. And that would be helpful as more of a massive consumer market, if you like, so that everybody who's using it, you know, so I'll give you an example. In Google Docs, you can set it up as enterprise so that if I start typing certain things in there it will say you can't share that because they're no insensitive words.

28:39Like in our world, for example, with the Department of War, with the Pentagon and you go down across CMMC, Cyber Maturity Model Certification, which is the whole supply chain cyber scenario for defence industrial base. We have it set so that you can't, if you put certain terms in, like unclassified but sensitive, it'll pick it up. The same principle you can apply to the chat window So it protects you from doing something silly yourself like putting your social security number in or some of your medical data in or whatever it might be. Yeah. How is the market? I mean, this is a new world with AI.