Loading summary
Transcript103 lines
- [00:00]
Patrick Gray
Foreign and welcome to another edition of Snake Oilers, the podcast that we do here at Risky Business, where vendors come onto the show to pitch you their products. My name is Patrick Gray. Everyone you hear in a Snake Oilers edition paid to be here. This is a sponsored podcast and we're going to be talking to three different vendors today about all about what they do. First up we've got Pangea. That's P A N G E A. And what they make is a product that is designed to put some security controls and guardrails around AI applications, which is a big problem at the moment when you've got enterprises building hundreds of AI apps and you know, especially when they're customer facing, you've got prompt injection problems, you've got people tricking them into saying silly things or tricking them into offering products for a dollar, that sort of thing. So we'll be talking with Oliver Friedrichs, who is the co founder and CEO of Pangea, in just a moment. Then we're going to hear from Cosive, which is an Australian company. I actually know the guy who founded Cosive, Chris Horsley. He's been kicking around Australian infosec for a very long time. When I first met Chris, he was actually working at the Japanese cert, so there you go. But he's got a long history here in Australia and his business is a threat intelligence shop that does threat intelligence consulting. But they're now offering a product that's proving to be pretty popular, which is a hosted MISP server. So MISP is an open source threat intelligence platform and you know, people tend to spin it up, put it on a box under someone's desk and sort of forget they own it. It's not maintained properly, they're not getting the best value out of it. So Cosive, Chris's company has done the work to figure out how to make it play nice on aws and they offer Cloud MISP as a product and they also have some consulting around that to help people use it properly. So that's our second pitch and our final Snake Oiler this week is sysdig. And sysdig makes a runtime like Linux security product that's pretty popular. Alex Lawrence is the director of Cloud Security Strategy at sysdig and he's joining us this week to just sort of pitch the product generally for those who aren't aware of it, and also talk about how they're using AI in their product to make it better. So that is a fun one as well. But we're going to kick it off now with Oliver Friedrichs over at Pangea. Pangea is a startup that's been around for a few years now, and what it does is put guardrails around AI applications, which is something that everybody kind of needs at this point, especially when they're, you know, doing things with customer facing AI agents. So I'll just drop you straight into the pitch now with Oliver Friedrichs from Pangea. Enjoy.
- [02:57]
Oliver Friedrichs
Pangea really builds the industry's broadest set of guardrails to secure AI. As enterprises increasingly deploy and build AI applications, some of the companies we talked to are building 900 gen AI applications. How do you secure that for your customers, employees, partners and so on? It's crucial that you protect against the latest threats. For genais, those are typically measured by the OWASP Open Worldwide Application Security project. They've classified 10 of the top threats. That's kind of the center of gravity for AI security. Today we protect against eight of the top 10. So we help you build, deliver and ship secure AI apps fast.
- [03:39]
Patrick Gray
Okay, so we're talking about this OWASP list. I've skimmed it previously, but it covers stuff like you would expect. Like things like prompt injection, right?
- [03:48]
Oliver Friedrichs
Correct. Yeah, that's really number one for a reason, because that's the main thing that people are concerned about. Prompt injection is essentially manipulating the application or model into doing something that goes against its basic instructions, like what the developer or administrator told the app to do via the system. Prompt. Prompt injection allows you to manipulate the model in a way that it evades that. So, for example, if you were told to be a pleasant support agent and I use prompt injection, I could teach you how to be a violent, horrible support person, for example, using profanity. You would not want that in an enterprise architecture, Enterprise deployment. Right? So that's number one is prompt injection. So we provide a prompt injection detection service to be able to prevent, detect and avoid prompt injection by over 99% accuracy.
- [04:39]
Patrick Gray
But that is a hard problem to solve. Right? And that's one of the reasons I was really interested to get you guys into this podcast is to talk about that. Do you? Because, you know, it's not like you can just have a banned list of words, right? Like you've got to be able to somehow infer the intent of the prompt. And the only way I can think to do that is by using another LLM. So I'm really curious to see how you tackled that problem. And of course we're not going to just get bogged down into talking 12 minutes about prompt injection, but I did want to ask you that, yeah.
- [05:08]
Oliver Friedrichs
I mean, look, this is, this is a challenging problem right there. There's no 100% solution today. You know, my background is in the anti malware space. I worked at McAfee in the late 90s, Symantec in the early 2000s, and this is eerily similar to that space. We're talking about words now instead of bytes. Right. Back then, we had to create detection logic using machine learning in many cases, to detect new strains of malware. We're doing the same thing now to detect prompt injection. And the interesting thing is every single day there's a new attack. Right. We've classified over 170 different methods of prompt injections so far. And we're building a very robust taxonomy with a group of PhD level researchers that work here that are focused on this problem. And it's fascinating to see how many different ways that you can manipulate large language models into doing things that you wouldn't expect them to do. For example, did you know they could talk in Morse code or in Caesar cipher? Right. You can almost instruct them to do these radical, interesting things. So how do you prevent that? So to your point, we need to actually leverage large language models and Gen AI to detect prompt injection because that's the only way to actually determine whether the output from a model matches the system intent and the system prompt that was issued originally. So we're seeing these attacks evolve almost daily where we need to respond and retrain our models almost every single day and issue updates to our customers to protect against the latest evolving prompt injection attack. So it's definitely an interesting space and evolving very, very quickly.
- [06:40]
Patrick Gray
Well, I mean, that's the thing, right? If you're offering a product in this space, it doesn't need to be 100% accurate just yet because currently the alternative is to use nothing, and that ain't good. Right. So I'm guessing what's bringing in a lot of your early sales, right? Because I understand you're a fairly new company. What's bringing a lot of this, a lot of the sales in is just would be people who just need to put something there to prevent trivial prompt injection from doing weird things to their agents.
- [07:07]
Oliver Friedrichs
Yeah, like a great, a great example. One of our customers, Grand Canyon Education, you know, they built a chatbot for their students, for their teachers to be able to provide support to that community. And they wanted to prevent PII and confidential information from leaking right out of their chatbot. Right. So that's where they use Pangea's redact service to be able to prevent that data leakage. So that's a very low hanging fruit.
- [07:32]
Patrick Gray
Okay, so that's a different use case which is more around model output rather than model input. So I'm guessing that would be. Yeah, that would be a big one as well. Right. Which is to just have that guardrail on there which says if you start seeing this model coughing up people's Social Security numbers, maybe get it to stop.
- [07:47]
Oliver Friedrichs
Right, Correct. Yeah. And you'd be surprised at how many different tricks you can employ to actually get models to emit data like that. Different encoding mechanisms, you know, use the first letter of every word to encode a message. Right. Those are all tricks that you can employ with large language models. So being able to detect those attacks both on the prompt coming in, but then detecting them. For example, let's say the actual prompt injection attack succeeds, you still want to capture and detect and block it on the way out through the output from the model as well.
- [08:19]
Patrick Gray
Yeah. So I'm guessing that's those are going to be your sort of primary use cases, is stopping people from doing, you know, easy to do weird prompt prompt injection attacks and then monitoring the output. Or have I got that wrong? Is there some other killer use case that people are just like, oh, you do that solved.
- [08:34]
Oliver Friedrichs
You know, I think there's really four categories. The first is prompt injection.
- [08:38]
Patrick Gray
Right.
- [08:38]
Oliver Friedrichs
So that's the one we've been talking about. Very complex and evolving very rapidly. The second is malicious content. So you definitely don't want malicious URLs, malicious domain names, malicious IPs, or other content being entered in through the prompt or through a mechanism that we call indirect prompt injection, where you're actually training data or using it for rag.
- [08:57]
Patrick Gray
Yeah, you give it a URL, say ghost scan this URL, read this page, and the prompt injection's there, right?
- [09:02]
Oliver Friedrichs
Correct. Or the model could emit it as well. Like, if you think about it, these models have been trained on the entirety of the Internet and human knowledge, which includes a lot of good information but a lot of garbage. Right. So it's encoded in there. How do you prevent that from coming out in an enterprise use case? Like for consumer use cases, you may actually want to know everything, but for enterprise commercial use cases, you have to protect yourself. So malicious content is really number two. The third is really the confidential information PII that I mentioned, being able to detect over 50 types of PII and filter that or apply what we call format preserving encryption to that data. So it still looks like a Social Security number. But it's now encrypted so that only someone with the right access can see it. And then you have other filters, for example, like toxic language, profanity, violence, self harm and so on. Or even competitive language. Right. If you're a car manufacturer, do you want someone asking about your competitor? Right. Absolutely not. Right. So there's, there's a lot of guardrails that we have to implement around this to be able to prevent things that you would normally expect in a consumer environment from happening in an enterprise use case.
- [10:06]
Patrick Gray
Yeah. So right at the top, you said, you know, enterprises are building like some of them, like 900, you know, they'll have 900 models that they're using to do various things. But I'm curious where the uptake is for something like this is. I mean, the reason I know that you're a new business is that these types of agents are actually quite new. Right. So you know, who's buying these guardrails for this AI stuff? Because, I mean, you're seeing LLMs pop up everywhere. Right. But I just wondered if there are particular verticals, I would imagine anywhere where they're using AI to interface with the public, that would be number one. Like, so anyone who's operating a decent enough support function through AI.
- [10:47]
Oliver Friedrichs
Exactly. That's really number one. Right. Is a support chatbot that offsets the need to have thousands of support agents. Right. Either locally or internationally. That's probably the first use case that we're seeing predominantly where an end user can interface with that chatbot to potentially manipulate it and divert it. So that's probably number one. In other cases, we see internal use cases for manning enterprise data as well. So typically what you'll see is that's where RAG is being used, typically sourcing thousands, hundreds of thousands, or even millions of documents, storing those in a vector database to be able to then combine that enterprise knowledge with the large language model. That's where you start getting real value from enterprise level data. Because these models don't know anything about your company. Right. They know everything about the Internet or Reddit or Slashdot or other sites, but they don't necessarily know anything about your particular business. So that's where other risks get introduced based on that data that's being sourced into that RAG pipeline.
- [11:47]
Patrick Gray
And that's where people really want to start watching the output of those models, I'm guessing.
- [11:51]
Oliver Friedrichs
Correct. And that actually introduces like a secondary risk, which is, you know, if I'm Oliver and I'm in engineering, should I be able to ask questions about Finance or hr?
- [12:00]
Patrick Gray
Yeah.
- [12:00]
Oliver Friedrichs
The answer is obviously no. But if you're dumping everything into a vector database, how do you provide granular authorization at a RAG chunk level, which is basically how documents are broken up into when you source them via a RAG pipeline, how do I honor the actual permissions of those original documents based on the identity of the user issuing the prompt? Right. So that's where applying authorization at a chunk level over RAG becomes crucial. And that's another service that we offer authorization in a way that allows you to map my identity down to the source documents being sourced via RAG and then honor those permissions. So I could only ask questions with my prompt on data that I'm allowed to access. And then more importantly, we have some open source libraries called multipaths that allow you to connect back to the origin file store and actually validate those permissions in real time during inference, so that you could determine even if the permissions changed, let's say on a Google Drive doc, that I could still access that document at the point in time where I ask a question as well. So we've built a lot of infrastructure. You know, while we sound new, we've actually been building the company for three years. You know, I'd like to say we. Oliver envisioned this.
- [13:12]
Patrick Gray
Oliver, three years is new.
- [13:15]
Oliver Friedrichs
It is new. Yeah. In AI world, it's, it's, it's ancient.
- [13:19]
Patrick Gray
You're dinosaurs. Yeah, no, I get it, I get it. Now look, one other question I have, right, Is that quite often in the wider cybersecurity space, a reason that people will buy a product is because they've experienced an incident, right? So I'm guessing that probably some of the customers who've come and bought your product have had some like somewhat hilarious horror stories. I just wondered if you could share a couple with us because I'm sure they're very funny.
- [13:45]
Oliver Friedrichs
Yeah, I mean, look, there's one example, and this is public. They're not a customer, but a car dealership in California had a chatbot. This is early on in the Genai world, right? And a customer was interfacing with it and they tricked it into selling them a car for a dollar and appending the language with a statement saying this is legally binding. So at that point, do you take that to court? Do you take the company, they didn't really get the car for a dollar. But those are the types of incidents that can lead to harm here. Nevermind the malicious content. A chatbot, again, emitting violent content, language, self harm and other dangerous language that becomes a liability issue and a potential legal issue as well.
- [14:32]
Patrick Gray
Yeah. But I guess my question is really what I'm trying to understand is whether or not people are being prudent and rolling this before something bad has happened, or whether or not they're hitting some of these issues first and then looking for solutions.
- [14:43]
Oliver Friedrichs
You know, the interesting thing is. Yeah, this is where an example, it reminds me of the Internet in the 90s, right, where we were building things so fast that security was the afterthought. That's happening again. Right. So I actually see a world where we're going to make a lot of mistakes before we actually implement guardrails. Now the fortunate thing is there's a lot of companies that are not new to AI or at least machine learning in the traditional sense. They've been using ML decision trees and other algorithms for decades in financial services, for example. So they already have a fairly mature process and compliance model around releasing these type of non deterministic algorithms, which is fortunate in that industry. But then there's other industries that have never used AI before. And that's where we're seeing a lot of interesting development, in particular with agents. Right. We've all heard about agents now that's sort of the future. This is the year of the agent where language models, large language models are being hooked up to code that can now execute tools and run a sequence of commands using chain of thought and planning in advance to know what those tools and what that, that, that pattern should look like. And that's really where you start introducing even more risk because now you have a large language model that's already non deterministic, trained on potentially risky data, telling you what commands to run with which parameters to run and what tools to execute. What could possibly go wrong?
- [16:10]
Patrick Gray
Oh yeah, I mean this gets, yeah. Once you start doing instrumentation stuff. This came up in a conversation I had recently with Chris Krebs and Alex Starmos talking about Deep Seq actually, and you know, about how people sort of misunderstand the risks. But like if you were worried about this model turning on you, when you would worry is when you start plumbing it through so that it can instrument various machines and run commands and stuff like that. So yeah, that's always, that's going to be a fun one to talk about in a few years when attackers actually start using it. But Oliver Friedrichs, thank you so much for joining me. That was very interesting actually, and I enjoyed that very much. All the best with it.
- [16:45]
Oliver Friedrichs
Thank you very much.
- [16:47]
Patrick Gray
That was Oliver Friedrichs from Pangea there. Now they're prepared to put their money where their mouth is. They are offering a $10,000 prize as part of their like AI escape room challenge. So if you just Google for Pangea, which is P A N G E A and Escape room, you will find it. And I think the URL is Pangea Cloud. So yeah, that was a fun one. I admit being a bit skeptical going into that interview, but yeah, it was, it was good stuff. Now it's time to speak with Chris Horsley who runs a company here in Australia called Cosive. And Cosive is a threat intelligence shop. They do threat intelligence consulting, consulting and whatnot. And they've launched a product recently which is cloud misp. Right. So MISP is like Stix Taxi. It's like a, it's, it's a threat intelligence platform, it's open source and it's a bit fiddly to use to maintain and to really get value out of. So what Kosive has done is they are now offering hosted misp, like cloud misp, along with a bunch of services to help people figure out how to actually get decent value out of threat intelligence. And, you know, just the hosted MISP is turning out to be really popular and not just in Australia. This is something that they're offering globally. So I'll drop you in here where Chris Horsley explains basically what MISP is. And you know what? They do enjoy misp.
- [18:14]
Chris Horsley
Very popular in both SOCKS and threat intelligence teams. Open source software for sharing threat intelligence. And then some people get very caught up on that definition of threat intelligence because for some people it's purely sharing indicators of compromise. So we're Talking about hashes, URLs, domains, and they absolutely have their place. But MISP can do more than that. Where we're sending reports which might be about threat actors or campaigns or vulnerabilities. There's a lot of capability in the MISP data model and some people take full advantage of that. Others are content with just These streams of IPs and domains that they're using for blocking at firewalls, for doing detection work. So very common to take your misp, integrate it with Splunk, Sentinel, your XOR or your Soar platform. So these are the very common use cases for it.
- [19:05]
Patrick Gray
So this is like an open source platform that's built for handling CTI data, basically is what, is what MISP is. Right, okay, got it, yeah.
- [19:12]
Chris Horsley
And then the other key bit of misp, I think, and why it has achieved a lot of success is it's got this big network Effect because if you know someone that's offering you a MISP feed or is running a MISP server, you can connect your MISP to their MISP and now you're receiving a stream of what they've got, you can send sightings back to them to say like, hey, we got that, there's a mechanism MISP to say, okay, we got your report. We don't agree that this domain is malicious. We think this is legitimate infrastructure. I'm going to amend your report and then the publisher can go like, yeah, actually you're right, I'm going to amend the report and republish. So this idea of the community, this threat sharing community, which could be one on one, it could be dozens of organizations, it could be hub and spoke model, there's different ways to configure these things, but it's a community. Collaborating on these threats too is the other key thing for me.
- [20:02]
Patrick Gray
Yeah, so you came up with the idea to offer like managed MISP instances because you were trying to build out some of these sharing communities and realised pretty quickly that this was not a. I mean, it's an open source platform, right, but it wasn't typically well managed in a lot of places. Was that about the long and short of it?
- [20:22]
Chris Horsley
Yeah, exactly. So, yeah, we were working on a national threat sharing platform and even when we were offering scripts and the like to help people set this thing up, it was still too hard for most CTI and SOC teams. So then we really had the idea, what if we take that pain away? People just want to use MISP. They don't want to wrestle with the 12 to 20 MISP releases a year and quality assuring them and monitoring and backing them up and working out how to plug it into their wafs and their firewalls and making allow rules and getting it through network architects. So if we get rid of all of that, people can actually get down to using the platform for what they want, which is plugging it into firewalls and siemes and threat sharing and all the rest of it. So yeah, the tooling just, just allows for this. It's the technical problem with threat sharing. Typically the problem is more the, you know, what are we sharing? Do we have time to draft these things? So it gets rid of all the engineering problems and just leaves you with the how do we share threat and talk? What are we sharing? Which is enough of an interesting challenge in its own right.
- [21:24]
Patrick Gray
Yeah, so we were talking earlier, before we got recording, and you said like quite often MISP was just spun up by someone kind of informally in the SoC, on a box, under a desk kind of, kind of vibe.
- [21:35]
Chris Horsley
Right, that's it. And a lot of the people who come to us and they're interested in cloud misp, they already know what MISP is and they know what it's used for, they know they like it and then they just want to make the pain of maintenance stop because they've started with this, you know, literally it might be running on a laptop beside somebody's regular workstation and they're going, okay, we want to do things but it needs to run stably now. We need to have like a proper production grade deployment and they just don't have the time, all the engineering to do that.
- [22:06]
Patrick Gray
Right. So I'm guessing this has been around. You've been selling this for a little while. Like how popular is it? Who's buying it?
- [22:12]
Chris Horsley
So yeah, we've been doing this for a couple of years now. We get inquiries from yeah, all sorts of places, finance, telecoms, resources, education. So there's a lot of MISP sharing communities out there and there's some big name ones for sure. So you've got places like FSISAC would be one and interesting. Go and look at their site. So they do, you know, Stix Taxi as one threat sharing model and they do MISP as another threat sharing model. But there's a lot, how would you say so that come out of the woodwork and there might be a collection of universities that, because these universities in this area have a close relationship, hey, let's share threats. Because we're all facing the same stuff. So yeah, it's really interesting to see these organic threat sharing communities that are around the place, but not advertise anywhere necessarily.
- [23:01]
Patrick Gray
Yeah. Right. So I'm guessing you just spin it up in an instance in the cloud, manage the patching, all of that sort of stuff. And that's basically the pitch here.
- [23:08]
Chris Horsley
Yeah, that's it. And we sort of re architected MISP in a sense so that it uses AWS sort of native platform, native features. So we get the advantage of all the HA stuff and the backup stuff and all the best patterns that AWS gives you. So it's not just like VM running on anything.
- [23:26]
Patrick Gray
Yeah, right, right. So you're not just yeeting like a docker container into some Kubernetes cluster somewhere like this is actually you've figured out how to make it work with AWS nicely.
- [23:36]
Chris Horsley
Yeah. And we spent quite a bit of time upfront sort of engineering this to run the way we want to do and to take advantage of as much of the AWS sort of feature set as we could. And then it's all the ongoing operations after that. So, of course, you know, it's the monitoring, it's the upgrades. The other big part for us is just the support not only of, you know, is the platform running stably and, you know, you hit a bug or whatever it is.
- [24:00]
Patrick Gray
Let me guess, let me guess. How do I do this?
- [24:03]
Chris Horsley
How do I do this?
- [24:04]
Patrick Gray
Yeah, yeah, right. Okay, walk us through that. Like, where are people struggling with that and like, you know, how do you help them through it?
- [24:10]
Chris Horsley
Yeah. So the classic is that MISP is a very powerful platform, but also has a lot of knobs and dials.
- [24:15]
Patrick Gray
Yeah.
- [24:15]
Chris Horsley
So quite often the team knows, like, okay, so what we want to do is we want to take these intel reports and send them over this way, but not these intel reports. They are purely internal. And then there's another set of reports, again, that goes off to a different audience, but that should be unidirectional, like, they don't need to push back to us, we're just going to publish to them. So what are the patterns to do all of these things? And there are patterns for all of this, but this is where we can say, well, yeah, like, here's the menu of options you've got, and based on what you're telling us, option two and four, that's what you want to go for here. And we save a lot of time from people doing a proof of concept and sort of feeling their way through all this, because we've done a lot of this before.
- [24:54]
Patrick Gray
Now, look, you know, you're a longtime listener of the show. You know, I like to dunk on the CTI people, but. But I also recognise that it's a fact of life that people, especially in large organizations, are going to be using a tech like this. Right. They need to be looking for these IOCs. If they're popping up in logs and whatever, it's just something you've got to do. Right. But I guess one of my questions would be is what's the general state of this stuff out there? Who's using it? How common is it? Is it growing? You know, I'm guessing from what you're telling me that it is. If people. If there is a market need for people to actually buy hosted misp, I'm guessing that it's a growth area. But, you know, can you just give us a bit of a rundown on, like, you know, what's going on out there in CTI land with MISP in particular.
- [25:41]
Chris Horsley
Yeah. So, I mean, it's really interesting to look at the last 20 years because that's how long I've been doing cybersecurity now. And that was before it was called cybersecurity. And cyber threat intelligence wasn't even really a term of art. So that tells you a lot. And we often call it just data sharing between national certs. And these were CSV files and everyone had their own bespoke formats. And then you had to write a Perl script or a bash script to parse that. Org's format. So what's happened in that last 20 years is that we've seen this emergence of standards and MISP is one and STIX is another say. So now we've got at least sort of some commonality, so you can do a bit more plug and play that they're publishing this feed in those formats. I get the right tool, I can ingest, I can do stuff with it. So we started Cosive 2015 and we started as like a big part of it was being a CTI shop, you know, not publishing cti. We were much more interested in the tooling and the practice of cti. And we think in retrospect we were just a few years too early because in the last almost 10 years we've been running Cosive, you've seen a lot more organizations know what CTI is. They now have an idea of what to do with it. I think back in those days, the idea of threat intelligence was so nebulous that, like, what is it? For some people it was like IPs I want to block. And there's a whole debate about pyramids of pain and like, what's the value of just blocking IPs is that even threat intelligence? And you can talk about, you know, the context. And I think more and more people are understanding what you do with it. So something else we're really excited about, late last year there was released a CTI CMM Capability Maturity Model. There's sort of similar things been around for longer with, you know, SOCKS and security operations centers, but this is one we really like. For the whole idea of defining what should I do with threat intelligence? You know, we've got three tiers of maturity, so you can start very simply with, I'm going to pick these domains and these capabilities. It's very good at talking about stakeholder engagement and what are my intelligence products. Because the number one thing we've seen where the best intention CTI programs go to die is where we buy the tools and we buy the feeds and we buy the analysts. And then we start to answer the question of like, okay, so who's in our organization is going to do something with this and what do they want, what do they need, what format, what are they going to do? So it really hammers that point of like, before you start any of this, what are we, the intel team, going to offer as services to our organization? So half of that is requirements, like, what are we tracking? What do we care about as an organization to be the eyes and the ears of that organization? And then the other half is the intel products. So are we producing, you know, your classic PDF style reports or reports on the wiki? Are we doing IOC streams to the SoC? Are we helping the hunting team with looking for new techniques informed by, you know, what we're getting from some of the best intel providers?
- [28:32]
Patrick Gray
You mean there's a model that says people should figure out what they're trying to do before they do it?
- [28:35]
Chris Horsley
This is.
- [28:36]
Patrick Gray
It seems sensible. I gotta be honest, when you say.
- [28:39]
Chris Horsley
That, you're going like, well, of course, but we can get. Really. Because I, my theory is a lot of CTI analysts come from a technology background and they tend to be very technology first about things. And when I meet CTI analysts who came from maybe military background or some other background, they kind of understand like, what's the point of all this? And then it's like, okay, what tools do we need to accomplish that goal? Rather than, hey, I've got this cool tip or I've got this mist to play with and now I better work out what to do with it. So it's all about coming at things in the right order for me. Yeah.
- [29:12]
Patrick Gray
And what are the different, like, maturity levels that you get in a model like this?
- [29:16]
Chris Horsley
So typically now from memory in the CTI CMM3 levels, and it kind of be what you expect. So at level one, we have some basic capability where maybe we're handling, you know, lists of indicators of compromise. It's that basic stuff and we can put it in to our block list and into our seam and then all the way going through to, you know, tier three. Now we're doing things like generating intel and we're doing our own research. And this is where we have analysts in our team. I know here in Australia it's a rarity for a lot of organizations to have, and I don't want to say it's rare because there are organizations who have dedicated CTO analysts now. That's Been a big change in the last 10 years. Before it was a lot of, well, someone in the SOC just likes reading blogs. So they're kind of like the intel analyst for us. And that is, you know, in some ways that's, that's the fledgling intel capability of like, what's going on at that threat landscape sort of stuff. And then we build up to like, okay, now your full time job is doing that and you've got budget, you've got tools, you've got feeds, we've got some structure, we know what the outcomes are supposed to be and that's how you kind of move through this maturity.
- [30:25]
Patrick Gray
Right. And that's something that you're offering as a sort of consulting service along with the cloud MISP stuff. And not just in Australia either.
- [30:32]
Chris Horsley
Yeah, that's it. So cloud MIS in particular, we've seen really good international interest for this. And then alongside, you know, the operation of the instances themselves, it's helping people get to the point of like, we're getting utility out of this, we're getting value, we're getting use. How do we set this up, what are the outcomes? And helping people define, yeah, what do stakeholders want, what are our intel products, how do we deliver them? So this is all consulting. We're providing sort of either completely separate to MISP or alongside it in many cases.
- [31:01]
Patrick Gray
So there you go. If you want to spin up some cloud MISP and you don't want to run it yourself, you can reach out to Cosive. And Chris Horsley, great to see you again.
- [31:09]
Chris Horsley
Likewise, Pat.
- [31:10]
Patrick Gray
Always good to see you, Chris. Yeah, so as I said at the intro there, I've known Chris a long time from, from around the traps here in here in Oz. Thanks for filling us in on what it is you're up to. Appreciate it.
- [31:21]
Chris Horsley
It. Yeah, thanks very much, Pat. Much appreciated.
- [31:24]
Patrick Gray
That was Chris Horsley from Cosive there and you can find them@cosive.com that's C-O S I V E.com and yeah, good stuff. And it was great to see Chris as well because we've been bumping into each other at conferences for something like 20 years. It is time for our final snake oiler today and we're speaking with Alex Lawrence, who is the Director of Cloud Security Strategy at sysdig. I'll confess that this was a really fun interview. You know Alex, Alex is my, my type of people and, you know, you'll probably hear what I'm talking about as you listen to this one. So sysdig make a Linux security Agent, I guess, you know, it's a, it's a runtime security product for Linux. It's been around quite a while. People seem to really like it. So when they asked if they could come and do a Snake Oilers, yeah, I jumped at it. They've been on once before, but that was quite a long time ago. So Alex in this interview recaps what sysdig actually do and then he talks about some of the fancy stuff they're doing with AI. Right, because everyone's doing fancy stuff with AI now. So I'll drop you in here where Alex begins by explaining what sysdig actually is. Enjoy.
- [32:32]
Alex Lawrence
Sysdig is a runtime solution for, for security, right? We, we are built in a cloud native way. We are built with kubernetes and containers and the modern stack in mind. And the whole goal is to look at things in a real time context. So a lot of security tools will be focused on all sorts of things, right? Lots and lots of various different things. And it's kind of broken up into two major categories. Preventative controls, detective controls. Sysdig does preventative stuff. Everybody does preventative stuff. The thing that we're focused on the most though, is that detective side of the house. How do you deal with security in real time? When things are ephemeral as all get out, things show up and disappear in seconds and things change in seconds and then the environments are gigantic, right? So how, how do you do real time security when that is the environment you're trying to do something with? And that's the problem that we're aimed at solving.
- [33:28]
Patrick Gray
Okay, so where do you guys plug into the whole equation? You mentioned Kubernetes earlier. Is it this solely for Kubernetes sort of stuff or is it, does it work in all sorts of different places?
- [33:37]
Alex Lawrence
Places, yeah. So all sorts of different places. The little secret I always have said about sysdig is that anywhere you're running Linux we have value to bring you. At its infancy, sysdig was built to interrogate system calls. How do you basically speak the language of this new stack? What's the most, least common denominator of the way information is traded back and forth? And that is the system call. So if you think about the old days, what did you instrument? You instrumented your network, you instrumented the packet. You used Wireshark, right? You grabbed every single packet. You looked at, all your applications were doing. You could do really cool observability things, you could do really cool security things, you know, a la Snort. Other tools like that, that were out there. But once we shifted to the cloud and we stopped owning our data centers, we no longer own a switch. So what did you instrument, what do you plug into? You can't hit a span port, you can't replicate all of your packets, you can't interrogate them. You could do some fancy work with port replicas and some junk, but it was really, really complex. You basically lost that single source of truth that was the packet. So if you think about the cloud, what, what becomes the new packet and effectively it becomes the, the thing your applications are speaking at. So with a container or a Kubernetes app, or a thing running on like a Linux box, an EC2, whatever you might call it, that's a system call, right? Every single system call is how you're gaining access to resources, allocing memory, going and sending stuff out to a socket. It's all happening at that particular level. And those system calls, just like a packet, they don't lie. So if you can interrogate that system call, you can know every single thing happening on that host. And you could do that in real time? Yeah, in the cloud, it's like a cloud lock. So being able to go look at what your cloud objects are doing, you know what changed in rds, who logged in without mfa? You know, what are all those things that are going on in Kubernetes? It's the Kubernetes audit log. Right. All of these cloud services, they all have a thing that acts like that system call and that's your single source of truth to do real time security.
- [35:41]
Patrick Gray
Right. So I'm guessing that with the sysdig you think of it, everyone likes to talk about how they're a platform, right. So we'll just call it a platform for now. So I'm guessing you look at things like various log sources and whatever, but I'm guessing you also shim in what, some sort of kernel extension or whatever to collect syscall information.
- [35:58]
Alex Lawrence
Yeah, these days it's ebpf of course. Yeah. In the olden days, quote unquote, I've been here six and a half years, it was just a kernel module, right. But now with the advent of what kernel? 4.12 and newer, you got this EBPF extension that lets you do stuff in a much safer way. So that's the preferred path. We do still have some customers out there who I think are still running RHEL 5, God bless them. And in those cases we still use kernel modules, but in modern architectures we're EBPF these days.
- [36:25]
Patrick Gray
So I'm guessing how this works is, you know, it's essentially an agent that gets shimmed in automatically in environments where your presence is part of this sort of, you know, part of the process of spinning up new, you know, kernels.
- [36:37]
Alex Lawrence
Yeah, exactly. Some people bake it in as part of their image so that when they deploy the host, it's already there automatically. Some people use deployment mechanisms, you know, be that any insert random DevOps tool of your choice. Some people will do it in obviously our sweet spot of kubernetes. And so then it's just a daemon set that goes and deploys the agent across the nodes. So there's about 1001 different ways to do it in the modern age. But yeah, effectively it's sticking an agent on a thing and then being able to go and see all the stuff coming from that thing.
- [37:06]
Patrick Gray
Yeah. And what sort of stuff are you likely to catch with sysdig? Right, like what sort of attacks, you know, what sort of odd events are likely to get flagged by this. I'm going to use a word that you're not going to like by this agent.
- [37:18]
Alex Lawrence
I personally don't mind calling it an agent. You know, that's industry nomenclature.
- [37:22]
Patrick Gray
I'm glad we can call it an agent. So what sort of stuff is your agent likely to prevent, to catch, to detect?
- [37:27]
Alex Lawrence
Yeah, so like snort, it's kind of whatever your creativity limits you to. And so what I mean by that is it's interrogating system calls. And so that's everything happening on the host. So that could be like a shell being opened up. It could be someone spawning the netcat process. It could be a actor doing a chamat or a system call as opposed to running the process. Right. It's. It's anything that's traversing that host, asking for resource. And so you can catch all sorts of crazy things and it's really up to how creative you can get. I've worked with some people who want to log every single time a file is opened or closed or processes executed. Every time a packet is sent, a socket is touched, like, you know, everything happening on that host. They were producing a ridiculous amount of data and their SIEM must love that them because that's gigabytes and gigabytes a day. Right.
- [38:27]
Patrick Gray
Build per line of ingest. Someone's popular.
- [38:29]
Alex Lawrence
Yeah, Right. I've got some who want to go a little more abstract, who are really caring about very specific use cases. Right. Like they want to look for container escapes or they only want to look at it. If it's this and that and else this, you know, it gets into some fairly specific nuance of how they want to do the detections, but it really allows them to get very specific on what they're trying to accomplish.
- [38:53]
Patrick Gray
Yeah, I mean, does it pain you? You weren't pained when I said it was an agent. Does it pain you if they called it kind of like EDR for Linux?
- [39:01]
Alex Lawrence
No, I mean, honestly, like it can be used in that capacity. Right. Like that's. Yeah, that is an area where I would argue that people have ignored. Linux is great and amazing. I've built my career on top of it. But that doesn't mean that attacks don't still happen in that world. You don't need to have real time detections on top of Linux. It's saying that I've got a Mac, so I'm secure. It's not that it's inherently better or worse. They've got a different design paradigm, so it makes things different. Right. And the threats still exist. You still have to be able to tell when stuff is going on. And arguably Linux runs the Internet. Yeah. And so it is the target of choice when we start talking very large scale applications and things that we're doing these days.
- [39:44]
Patrick Gray
Well, it runs the Internet and it also happens to run a lot of coin miners, which I guess has been a big driver, a big driver of adoption. So you know, what sort of enterprises, what sort of organizations tend to, you know, tend to be running sysdig. And then, then I want to talk briefly about like how the products change because you've been around for a while, right? And this is always a moving target, you know, running a product like this. So they don't want to hear about what's new with sysdig. But where does it tend to pop up mostly? I'm guessing it's mostly, you know, anyone who's doing like DevOps style stuff, which is I guess not really modern anymore. I was about to refer to it as modern, but then again I am talking to a guy who keeps referring to snorts, so you know what I meant.
- [40:24]
Alex Lawrence
Yeah. Now it feels modern to me. Right. But yeah, no, DevOps is just like, that's status quo these days. It's not a new thing.
- [40:31]
Patrick Gray
Yeah. It's just the way it's done. So. But I mean, is that where it sort of tends to pop up is people who are, you know, running their own applications in the cloud and whatnot? I'm guessing that's where it's most popular, right? Yeah.
- [40:42]
Alex Lawrence
I mean the sweet spot for Sysdig is effectively anybody who is doing kubernetes and containers. Right? That's. That is the thing that we do the best out there. I'm not going to say that other folks don't attempt or try or do things, but we do put a lot of effort behind the way we do detections in that world. In particularly the way you do policy and enrichment and the way you can kind of handle that is a little bit more mature in the Systick model. And then a lot of that shockingly tends to be Finserve customers as well. Like Finserve is pretty darn progressive when it comes to this cloud native era of things.
- [41:17]
Patrick Gray
Well, they are because they're running really important applications in the cloud, you know.
- [41:22]
Alex Lawrence
Right.
- [41:22]
Patrick Gray
That mobile banking app or that brokerage account, you know, that is Linux in the background, right?
- [41:27]
Alex Lawrence
Yeah, well, and what's funny is that a lot of these organizations are running applications in cgroups and if you go back way long enough, you know, BSD jails things that fence off the processes so that you're not conflicting with other co running applications. That's all a container is. Right. It's just a big giant API ball around cgroups. And so this notion of containerization really isn't different for them. Right. Like they've been doing this for a long time now. There's just an actual standard they can follow.
- [41:55]
Patrick Gray
Yeah. So look, let's talk now about, you know, product evolution because everybody's shimming AI into everything now. I understand you're also doing a bit of this, like how do you start to apply AI to something like a Linux endpoint agent?
- [42:12]
Alex Lawrence
Yeah, that's a great question. It's hard.
- [42:15]
Patrick Gray
Yeah. Because I'm guessing most of the value you're covering off pretty well right. On the endpoint agent in terms of just being able to collect that telemetry. I'm guessing that where you would apply the ML, the AI, all of that magic pixie dust is going to be more on the correlation side when you're actually looking at the information you've collected off all of these endpoints and looking at other logs and trying to draw some insight there.
- [42:36]
Alex Lawrence
Yeah, no, it definitely is. Like, look, if we think about what is the problem AI is solving fundamentally, AI is addressing the data lake problem. Right. We've got a lot of data, more than we can ever do anything with, with a human being. So how do we do something effective with that? We use machines and in this case we used LLMs. We use AI this portfolio of products in the security world where, where we exist, it's this Gartner term, cnat, right? Cloud native application protection platform. It's a mouthful, but it generates data like no other, right? It's just as difficult to deal with the amount of information Cnapps produce as other security tools. You know, times 10, times 100. And so if AI is built to solve the data lake problem, it's actually uniquely positioned to help drive interesting insights into a cnapp tool. And that's the way we approached it from the beginning. We didn't write a thing that just read our documentation and told you something. We decided to teach it about our API and teach the LLM about how to interface with the way sysdig generates and visualizes data. And so yeah, we've got an AI integration. It's called Sage, right? It's your assistant for going through all of your systic information. But you can ask it things like on this page that I am looking at, what are the, you know, top two or three most important events that I see? And then of those events, what are they related to? What other events may have happened that I'm not saying that have come from this. And so basically it's helping you sift through that lake of information and doing it in a very pointed way, in a way that sysdig understands. Right? It's making API calls on your behalf.
- [44:31]
Patrick Gray
I mean, that seems like a, like a sensible thing to do. I do wonder though, because it seems like a lot of the sort of SIEM companies and whatnot are making tools to kind of do that on the Siemens. And I imagine you're already pumping a lot of this telemetry off to the Siemens. So I guess we sort of are at that point where we're working out where the AI best plugs in, right? Because I can also foresee be a circumstance where you do some sort of LLM processing on alerts to help figure out what to send to the SIEM to be processed by their LLM. It's like a little chain of like robot workers who figure it all out, right? I mean, we are still working all of this out, right?
- [45:05]
Alex Lawrence
Oh, we certainly are. And honestly, if the SIEM vendors weren't investigating what they can be doing with an AI agent or an LLM, you know, on their stuff, they're missing the boat, right? Again, AI stuff today is solving data lake problems, and the SIEM is exactly that.
- [45:22]
Patrick Gray
And what's the uptake been like, is this in beta or is it already out there? And Sort of what's the response been like from customers? Because you know, you're making a Linux tool, man. You're dealing with like crusty people. Right. So what do they make of this newfangled LLM enabled sysdig old school tool with new school tricks. Right.
- [45:39]
Alex Lawrence
I get them. They, I grew up as them.
- [45:41]
Patrick Gray
Yeah.
- [45:42]
Alex Lawrence
So it is getting decent adoption. Right. Like I've seen, we, I think we've seen 300 some plus percent growth across our current user base. Right. They were very tentative at first. Right. Like what, what does this mean? Should we be touching this? Should be using it. And we've, we've seen decent uptick in adoption of using the service we've been building out. And a lot of it comes down to sifting through events quickly. If you think about the way cloud attacks work today, again, especially on this cloud native infrastructure, being able to sift through data fast is basically your, your advantage.
- [46:17]
Patrick Gray
Yeah. Being able to ask natural language questions about a data set, I mean that's always going to be popular, right?
- [46:24]
Alex Lawrence
Yeah, exactly. Like I think the statistic from our usage report we put out every single year is containers are basically just getting. Their shelf life is getting smaller and smaller and smaller. I think as of the report that just came out this year, like just, just like a month ago or so, 60% of all containers live less than 60 seconds.
- [46:45]
Patrick Gray
So it means you can't just use GREP anymore. So we've spent as a society hundreds of billions of dollars to basically get to where we were with GREP 15 years ago, which is a funny old. It's a funny old world. We're going to wrap it up there. Alex Lawrence, thank you so much for joining me on the show to walk through. Yes, sysdig and what you're doing with AI. And just to recap for everybody, you might not know what it is you do. Pleasure to chat to you.
- [47:09]
Alex Lawrence
Yep, thanks so much.
- [47:11]
Patrick Gray
That was Alex Lawrence from SYSDIG there. Big thanks to him for that and big thanks to sysdig for being one of our Snake Oilers this time around. That's actually it for this edition. We'll be back with part two of this round of Snake Oilers, which will be three more vendors. We'll be back with that in a couple of weeks and in the meantime we'll be publishing as usual. So yeah, catch you all soon. Until then, I've been Patrick Gray, thanks for.