EP119 RSA 2023 - What We Saw, What We Learned, and What We're Excited About

Security.

Hi there. Welcome to the Cloud Security Podcast by Google. Thanks for joining us today. Your hosts here are myself, Timothy Peacock, the Senior Product Manager for Threat Detection here at Google Cloud, and Anton Chuvakin, a reformed analyst and senior staff in Google Cloud's Office of the CISO. You can find and subscribe to this podcast wherever you get your podcasts, as well as at our website, cloud.google.com/podcasts. If you enjoy our content and want it delivered to you piping hot every Monday, please do hit that subscribe button. You can follow the show and argue with us and the rest of our Cloud Security Podcast listeners on LinkedIn. Today's episode, very fun. We are talking about RSA and some of the awesome AI stuff that Google launched. But before we get into that, Anton and I, we're going to just spend some time talking about what we saw for those who were there or weren't there. I think you'll find this interesting. So Anton, what did we see that was notable?

I mean, I would probably start with AI. I mean, not because it was everywhere, but because frankly, I was looking all over multiple booths to see whether they're like oozing with AI.

They were not.

They were not.

But you have a theory about this.

I have a theory about this indeed, because I feel like those booth materials were printed, were manufactured before the current AI craze. So—

So we had a fossilized piece of security industry marketing.

That's exactly correct. That promises, that claims to be quick and responsive, but ultimately wasn't in this case.

It wasn't. And listeners, if you're wondering who's laughing, let me introduce today's guest. Today we're joined by Connie Fan, who's in our product operations and strategy group.

Is that right?

That's right. Yep.

Connie is joining us to talk about the AI stuff, and I think to join us about our nonsense talking about RSI.

I'm a sucker for medium jokes, apparently.

I think that's right. One of the things I thought was interesting was that we continued to see countries showing up with booths. We saw Spain, Korea, Germany, and Flanders.

Region of Belgium. And we were briefly arguing— North region of Belgium. Yes, correct. The north region, which I only managed to inform Tim of this because I read it on the booth faster.

So one thing we thought was missing was XDR. Where did it go?

Yes, I was preparing some kind of a pithy remark about this, but I feel like a couple of years ago, maybe last year, XDR, the extended detection and response, was everywhere. We're staring at you from every booth. Today, even a lot of—

had no business being.

That is correct. That is correct. And I think now we see some of the opposite. People are pivoting away from XDR. There was a vendor who became a SIEM.

That was interesting. Why do you think they became a SIEM? Because they finally have a Snowflake partnership?

No, I feel like you're trying to make me say something that isn't very critical of XDR, I guess. And I am going to be quiet about it because I don't have anything. To me, SIEM is still where the money is in terms of detection, response, monitoring. And it's a lot easier to do SIEM unless you're an EDR vendor. It's easy to do XDR. But for everybody else, if they want to have a broad-scope platform, it's probably a SIEM.

And I think that makes a lot of sense. SIEM is where the center of attention is. It's where the center of data gravity is. It's central to people's organization day to day. So I get why if you're a vendor, It's maybe more tempting to go after the SIEM budget than it is to go after the—

Mythical XDR budget. Correct.

Yeah.

Okay.

So on the note of DR, I saw TDIR. I assume that's not Timothy Dillon Incident Response.

No. And you can blame my former colleagues from the well-known analyst firm for that one.

I blame them for a lot.

You do, actually. And it's not always deserved because ultimately when Gartner coined the term TDIR for threat detection, investigation and response. I don't think they meant for it to become a market. I think they meant it to be a concept of operations at companies.

Is that like ASO as a concept of operations?

Kind of like ASO. Correct. Correct. Correct. But I think some vendors decided to stick it on the booth because why not?

I saw somebody stick a bunch of race cars on their booth and I never did figure out why they had all those race cars. So what about ZT? I think ZT was MIA and SASE seemed to be AWOL. Where, where were those this year?

I think I saw some zero trust. I mean, zero trust wasn't an oppressive presence in the air.

Not like it was last year.

I think last year was the year where it was everywhere as well, just like XDR.

You couldn't swing a dead firewall without hitting ZT last year.

You promised no firewall jokes today.

I was not— okay, fine.

So in any case, I feel like people who did zero trust network access, there were a couple of vendors promising ZTNA, and I think there was a SASE sneaking up somewhere. But ultimately, as we all know, it's a lot more difficult to implement than to paste it on the booth. So SASE and its younger brother SSE— don't ask.

I won't.

Weren't there in force, but I think some of the vendors who are relevant to this did cover zero trust.

Interesting. On the note of too many acronyms, listeners, I thought it was pretty interesting to see CISA placed so prominently, the Cybersecurity Infrastructure Security Administration, cyber— whatever the hell they are. The guys in the Department of Commerce that are led by They had a huge booth, very centrally placed in the South Hall this year.

I think it was bigger than NSA and FBI, maybe not combined, but it was big.

It was. NSA and FBI were next to each other this year, and I didn't see any interagency shooting matches break out. I also did see that the NSA returned with their Enigma machine, as they do every year. So listeners, if you do get to the conference in the future, I will say it's pretty cool to see an original Enigma machine. So one thing that we did see return in force was cloud security. Saw the 3 startups with big booths and one big player with a very small demo station. What did we learn visiting them?

So this was a fun project, I guess, for Tim and myself. We went across the RSA show floor looking for cloud security.

We had to. It's in the name of the podcast.

Right, exactly. So it's reasonable that we would be on a quest. And we sort of talked to some of the vendors who focus on this. Mostly, some of them are building larger platforms. And we knew that they weren't just doing posture management. They aren't just CSPM. But ultimately, we were sort of surprised how a lot of, I guess, their customers are still ultimately buying CSPM. They want improved posture management. So to me, the fact that most of the cloud security problems are ultimately misconfigurations, mistakes, vulnerabilities, and they do play a huge role. So even vendors who had detection components use detection in service of posture management, which struck me as interesting.

I thought that was very interesting. One of the vendors actually was pretty direct about it. And I was very impressed, by the way, those 3 vendors with the quality of the staff they had in their booths. So shout-out to those 3 vendors for having great people. He made it really clear that their users were so early in their cloud security journey that they weren't yet ready to consume the threat side of the house and really were just trying to get off the ground With the posture piece. I thought that was a very interesting, like, maturity capability thing going on.

Yes. Yet the other vendor mentioned that more mature clients buy a balance of CSPM and I guess CDR, but nobody said CDR.

Nobody said CDR.

Why didn't they say CDR?

I don't know why they didn't say CDR. You know what else they didn't say? They also didn't say CNAPP.

No, they did say CNAPP. One of them did say CNAPP.

It was the big one that said CNAPP.

By the way, if you know how to spell CNAPP, you're probably one of the maybe roughly 20 people on the planet, right?

It's got more letters than you think, listeners.

Yes.

Yes. They did spell it right on their booth. We did not visit our friends at Snyk. I think Anton might be afraid of dogs. We're going to look into that more. One of the things that I thought stuck out was the agentless versus agent-based approach. Do you want to talk about that a little bit? Because I thought that was pretty fascinating.

I think this is so '90s. I don't know. I'm not sure we need this debate on our podcast in 2023, but ultimately—

No, of course we do.

Okay.

It was central to the messaging from 2 of those 3 players. Actually, all 3 of those players were all about this pivot point.

Indeed. But it's the balance between having deeper visibility into activities on each workload kind of requires an agent. Each host, correct. Each host requires an agent. But then if you have an agent, then you're dealing with compatibility, other challenges. So, you sort of trade—

Resource consumption, DevSecOps to get it deployed, holes in the firewall to let it update itself. There's a lot that goes with an agent.

Right. So, you pay for depth with some complexity and other challenges, but ultimately, you do gain depth. And if somebody needs that depth, Agent-based it is.

But you gain depth only for host-based workloads.

For VMs, correct.

For VMs or GKE or Kubernetes clusters, but you don't get it for Cloud Functions. You don't get it for databases. You don't get it for any of the managed services. So, what's the role of the agent there? It doesn't seem they can have one.

I think it's a separate podcast episode. We've been meaning to do, but never did.

And I think it's time. Oh, good news. Separate episode. One thing that was also missing Was AI. Not a single one of them mentioned AI to us.

From these vendors, correct. But actually, one of them we know uses should we call Bard competitor? That's okay.

I'll allow it. But there are some unflattering comparisons that come to mind. How are they using it?

Oh, they're using it to make remediation instructions. If you paste an alert into the system, supposedly the big AI brain spits out what you should do.

Huh.

So Connie?

Yes.

I want to bring you in here because you've been working super closely on everything we've been doing in cloud security with AI. What do you think first about this use of generative AI to generate remediation steps? And 2, could you talk to us about what we're doing with AI?

Yeah, yeah.

I think it is an excellent use case. It certainly requires a lot of work, and we need to make sure folks are doing that work and not sharing dangerous hallucinations that are going to waste people's time. And we're very conscientious of that here at Google.

Could you explain real quick what a dangerous hallucination is?

As opposed to what, safe hallucination?

Yeah.

What's a hallucination, baby?

Depends on the substance you take.

No. That's the most San Francisco answer we've had all day. What's a hallucination in the context of a generative AI model?

Sure. A hallucination is basically when the model has made something up. The model is effectively really, really good at saying, if you've given me this sweep sequence, here are other things that make sense in the sequence. And sometimes it can tell you things that sort of make sense, but aren't really quite there. And when you really need precision, and when you really need to take the right action, you want the right recommendation, you don't want something that was kind of close, not quite there. And so that is, I think, actually on the safer side of hallucination. A dangerous hallucination would be when the model just tells you to do something bad.

Which ultimately increases your risk in the context of recommendations, right?

Exactly.

Ultimately, it tells you something plausible that increases your risk. That's like, I don't want this in my security AI for sure.

So that kind of scares me about remediation steps then. Like, what if it makes up the wrong command to run?

Or the command that is right and it would run, but would make you—

Less secure.

Yes.

Yeah.

Yeah, I think that's a good time to emphasize that as important as the models are, building the right systems around these models is also incredibly important. Setting up the right guardrails. If you are integrating these models into your products and you're actually going to be showing these to real customers who might not have the expertise to judge good from bad very well, making sure that you've taken every step you can to not show the bad in the first place is— Just actual work people need to do.

As a quick side note, you mentioned something, showing it to clients who don't have expertise, but ultimately the promise of several use cases around generative AI, as far as I understand, is to democratize some of the skills. So it is almost by definition and by charter, it would go to people who don't have the skills, otherwise people would just do the things.

Yeah, absolutely.

Absolutely.

And so in our democratization, because it's something that we're claiming too as well, right? In that, it's very important for us to understand what our customers struggle with and what are maybe not the flashiest skill sets. Like, we're not jumping right to remediation intentionally. We're doing things that are day-to-day tasks that are just difficult for people that we can make easier for them. And I think a great example of this is what our Chronicle team is doing when it comes to search. It's something that is so critical to every investigation, so critical to day-to-day tasks, And it might not be as flashy as what some of these other companies are claiming, but, you know, if you've ever had to write any query at all, Tim asks me what our revenue numbers are all the time.

I do, because I don't know how to query for them, and Tani does.

You know, you might be, you know what you wanna ask, you know why you wanna ask it, but the how can be really tricky. And knowing where that data lives, how to structure the query takes time.

It does.

Yeah. And so that's, In Chronicle, you can now tell Chronicle in natural language what the question is, whether it be how much money we're making, or if it is, you know, the example, I think if you read our blog post, is tell me about all the times confidential files were shared externally last week.

That's a hard thing to query for.

But it'll write that query for you.

That's a really cool example. And what I like about that is it's got a couple of safety aspects to it. You're not making changes with it. You're just making a search with it. And then you can see with your own eyes, your own lying eyes, whether you got what you hoped to get or not.

Exactly. And that's something that was really, really important to the design of this feature, where you can continue to iterate. The running thought is, you know, you'll see the query and you might even realize that the question that you started to ask Isn't actually really what you want to ask. And maybe you ran it and maybe you saw some of the results, but you'll keep refining. You can refine inline. The AI will suggest other paths to take as well.

It's—

but we never do anything directly for you. We're always still keeping you in the loop, still making sure you're comfortable with everything and that everything looks good to you as the user. It's just that a lot of time was saved. And in some cases, you might not have been able to do that answer completely on your own. So it fills in some of the gaps for you.

Yeah. But I think the generative examples are amazing, but explanatory examples, like the Code Insight with VT, it's like the whole concept of pasting some kind of obscure obfuscated script you found online on some malicious site into a tool, and the tool says, yeah, it does this, and it's probably from this threat group. That actually blows my mind actually more than the natural language search. I've seen natural language search. It's fine. But Anton's been around the block. He's seen it. But the malicious scripts and sort of like almost like reverse engineering badness without any skills and attributing it almost to the type of threat actor, that to me is quite amazing. So it's not only about what you can generate, but also kind of what AI explains, summarizes, classifies. That's actually kind of cool.

I think it's amazing what our VirusTotal team has done. Quick plug for them, actually, the first day, They made this feature available to customers. A new implant was uploaded with no antivirus detection. So they're truly, truly catching threats as they emerge. And to your point, it is, it almost feels like translation and summarization all in one. And it's just so indicative of all the capabilities these models have, or the wide range of capabilities, because it's translated from the code to actions. And instead of giving you a long, rambly response of, here are all the actions that were taken, it's—

Yep.

Here's the stuff you care about. I think it's so cool.

And given that it may be a programming language you don't even know at all, I can imagine the sequence of steps I would have to do to sort of understand what a malicious script does.

Imagine trying to read a Rust program.

Right. Like, you don't know Rust, but it's a malicious Rust program. That's a good point. And then in this case, the technology would tell you what it does and the type of maliciousness. That is actually amazing. And I mean, maybe I'm more impressed by this than with the generative examples, but I want my AI to write YARA rules and other detection and queries. But to me, some of the classification, summarization, explanation examples, like the whole write me a board report in the future when we get to write me a report about this type of a threat, that's also kind of cool. And it's also not the most favorite part of many security professionals' day job. That's true.

A lot of people would rather be Splunking the logs than typing the report for the board. I've been using AI in that classification context for a long time now in one of my products in production. With our container threat detection product, we classify Bash scripts in real time using an NLP model. It's not an LLM, but it is a natural language processing approach. And I've been really impressed with what we've been able to do there. We literally catch zero-day activity, not on the exploit step, but on the second stage step. And adversaries use Bash scripts in 90% of the honeypots we run. It's a very, very powerful opportunity. One of the things that one of the VT guys said to me about this effort was that, sure, it looks like a silver bullet today, But we haven't seen adversaries adapt to it yet. So I'm very interested to see, and I'm interested to see what our listeners come up with when it comes to adapting to this kind of detection.

This sounds like an RSA 2024 conversation, right? It does.

Maybe we'll have some interesting talks in the future. Connie, one of the other things I know we launched is attack path explanations. Could you explain the explanations?

Sure. Let me have a crack at it. Well, it sits in one of Tim's products, so maybe he's better off doing this.

No, no, no, no, no.

But effectively, SCC has, through very advanced modeling, already been able to simulate the path an attacker might take. But as you can imagine, for those who aren't incredibly well-versed in these steps, even interpreting that simulation and why they were compromised is still tedious and still difficult. And so on top of that already advanced modeling, we're layering in one more step of running everything through an LLM and being able to generate sort of like what we were saying with VT Code Insight of you don't want this long rambly thing. Not that that's what the simulations are today, but it's just hard for people to understand. A very quick synthesis of here are the key things that made you susceptible to compromise. And, and I think this is what the team did that's really cool. There you go. Here are the steps you need to take to remediate. And that is all generated by the LLM.

It's a really cool capability and gets back to that story around democratizing and making available security insights to people who don't want to read a 200-step attack graph.

And it's also a lot less fragile in regards to hallucinations.

Yes.

Because it's not like the system would write you code that then you have to run on your prod systems where hallucinations can be like, Lethal mistake. But in this case, it would probably be at worst a not very coherent explanation. I don't know what's the negative, what's the risk of this system doing a subpar job? Probably it's unlikely, but if it does, you're not going to die.

Right. I mean, listeners, we often have subpar explanations on the show and nobody's died yet. So Connie, one thing I think that's interesting is how close you were to this whole thing. And I watched you working on this for the past couple months. Could you share maybe just a couple of stories about how all this came together so quickly? How did the organization make this happen?

Oh gosh, yeah.

Listeners, there is just like a look of horror and terror washing over her face here.

That was my chuckle. If I watched Tim, then he was a very supportive friend and heard me be like, oh my gosh, this thing's happening. It's so exciting. Oh my gosh. All right. Well, I think the first First thing that we really need to emphasize is we sit within this amazing company that has been doing so much research for so many years and is really ready to productize that research. And that's why it can happen so fast for us, because we can hop on a call with the researchers who understand these models best and fine-tune the models and make it work for security use cases without Being within the organization that we're in, I don't think we would have been able to make it happen that fast. It's also a huge testament to our product teams who had roadmaps, who were stretched so thin already, and were able to think of— I mean, some of the coolest use cases are, well, we didn't think we'd be able to get to this until next year or even beyond that.

Mm-hmm.

But now, because of this new tool that's in our toolkit, They didn't change their mission anyway. It's just, we can table some other stuff and get to that hard problem so much faster. And that certainly wasn't a pain-free process for us to all have to break out of our traditional ways of thinking and the plan that we had. But yeah, it was a lot of early whiteboarding sessions that rapidly turned to execution. And that's That's how this all happened.

Yeah, it's remarkable to me how quickly this came together.

It's a useful reminder for everybody, well, outside of Google as well, that Google can move if need to.

Sometimes.

Sometimes. Fine, fine, fine, fine.

Connie, do you have any other thoughts you want to share on this? And I'm sorry to surprise you with this, but we always ask the same traditional closing questions of our guests, and they weren't in the doc. Do you first have one tip to improve how people are using AI? And 2, do you have recommended reading? Can be anything. Somebody recommended Mary Oliver last week.

Oh, fun. Okay. Wasn't the surprise question because that was assuming I've never listened to your podcast before.

Oh, you listen?

Uh-huh.

Uh-huh.

So recommendation of how do you say it? I imagine everyone at this point has played around with Bard, with Bard competitors. And the really nice thing is it makes it so accessible to people. Because before, I think the recommendation last year, 2 years ago, now is just get hands-on, see what works, what doesn't for you. And before, the easiest way to get hands-on would still require Colab, would still require basic Python knowledge. And now you just need to know how to talk English.

Exactly.

Wait, it talks in other languages too.

You don't actually need to know English.

Yeah, you're right. It's even better than that. And so you could do it in Russian, right, Anton?

I don't know. I haven't tried. Nope.

What? Well, we can verify that later.

Yeah, but still, it's just, it's get in there. See, when you slightly tweak the question, how does your response improve? Or how does the model's response improve? I mean, I'm going to misquote this, but there was recent research that said if you just enter the words explain You're thinking step by step to the model. It wasn't just that the answer you got back was more readable. It was that the answer you got back was actually more correct. And people can't explain why that is, but you just, to play around and find stuff out like that is so cool right now. It's so necessary. So everyone should just be hands-on.

Recommended reading, please.

Recommended reading.

It can be our podcast site. I don't know, something, you know, Google Cloud Blog. I don't know. I'm trying here.

I feel like everything that I'm reading right now is honestly other companies' blog posts even. And it's just so impressive what other— and yes, on the floor, maybe we could— other folks couldn't get printouts and demos out as fast, but there's a lot of really interesting stuff online right now of what people we're doing with AI. And I'm trying to think of if there's any specific blog. I honestly just think creating a news alert for generative AI security is giving a lot of good content right now.

Amazing. Well, Connie, thank you so much for joining us today.

Thank you for having me.

And now we are at time. Thank you very much for listening and, of course, for subscribing. You can find this podcast at Google Podcasts, Apple Podcasts, Spotify, or wherever else you get your podcasts. Also, you can find us at our website, cloud.withgoogle.com/cloudsecurity/podcast. Please subscribe so that you don't miss episodes. You can follow us on Twitter, twitter.com/cloudsecpodcast. Your hosts are also on Twitter at Anton_Chuvaikin and _TimPico. Tweet at us, email us, argue with us, and if you like or hate what we hear, we can invite you to the next episode. See you on the next Cloud Security Podcast episode.

Security!

Security!