Latest / Expert Insights Podcast / #66 – Zero Trust and AI in the Fight Against Phishing (Deepen Desai, Zscaler)
Transcript
- 0:03SPEAKER_01: Hello and welcome to the Expert Insights Podcast,
- 0:05SPEAKER_01: your go-to source for cybersecurity insights from industry experts.
- 0:09SPEAKER_01: I'm your host, Joel Wits.
- 0:14SPEAKER_01: Are we prepared for the rise of AI-driven cyber attacks?
- 0:18SPEAKER_01: In today's episode,
- 0:19SPEAKER_01: we explore this question and more with Deepan Desai,
- 0:22SPEAKER_01: the Chief Security Officer at Zscaler.
- 0:25We discussed how zero trust and AI are shaping the future of
- 0:28cybersecurity,
- 0:29SPEAKER_01: the growing risk of AI-powered phishing attacks,
- 0:32SPEAKER_01: and the fundamental importance of strong security frameworks.
- 0:36SPEAKER_01: Let's dive into it.
- 0:38SPEAKER_02: It's even great to chat to you again.
- 0:40One of the headlines for Z-Sketter this year at RSA,
- 0:43what have you been talking about the most?
- 0:45SPEAKER_00: Yeah, no, I had a presentation on Zero Trust and AI.
- 0:51How do you gear up to protect against the AI-driven attacks
- 0:58that we're starting to see?
- 0:59SPEAKER_00: Leveraging Zero Trust in AI, what does it look like?
- 1:02SPEAKER_00: So that's what my talk was all about.
- 1:04SPEAKER_00: I kind of gave a few examples as well.
- 1:06SPEAKER_00: One of them, you saw it probably at Zenith Live, the Rogue GPT demo.
- 1:10Then I gave examples of how ransomware operators are going
- 1:14after these AI vendors
- 1:15SPEAKER_00: as well,
- 1:15SPEAKER_00: because...
- 1:17One AI company getting popped means there's a huge downstream
- 1:20impact with multiple customers.
- 1:22SPEAKER_00: So in this case,
- 1:23they popped a company and they were able to steal unanonymized
- 1:28data of their
- 1:30SPEAKER_00: customers,
- 1:30SPEAKER_00: which were supposed to be used for training models by this AI company.
- 1:35SPEAKER_00: And they demanded $15 million in ransom.
- 1:37So I walked through that case study as well,
- 1:39like how we're going to see attacks evolve.
- 1:43SPEAKER_00: Using AI is one thing,
- 1:44SPEAKER_00: but AI being leveraged insecurely or your supply chain,
- 1:50SPEAKER_00: which will now also consist of a lot of these AI vendors,
- 1:54SPEAKER_00: will bring in those additional risks.
- 1:56SPEAKER_00: And we're starting to see examples of that across the board.
- 1:59SPEAKER_02: Yeah.
- 1:59SPEAKER_02: How have things changed in the AI space since we sort of spoke last year?
- 2:03SPEAKER_02: Has the threats accelerated, adoption accelerated?
- 2:06SPEAKER_00: I think it's all about agentic AI now, right?
- 2:08SPEAKER_00: I'm sure you guys are seeing that.
- 2:10SPEAKER_00: And we've heard a little bit about it.
- 2:11The definition of agents is getting diluted in some of the
- 2:16conversations that I'm hearing.
- 2:18SPEAKER_00: It's just a script.
- 2:19SPEAKER_00: They're calling it agent.
- 2:20But there is a lot of potential in the whole agentic AI
- 2:26workflow if done right.
- 2:28SPEAKER_00: But then again, it does come with its own set of risks.
- 2:31SPEAKER_00: We saw,
- 2:31SPEAKER_00: I don't know if you guys noticed,
- 2:33SPEAKER_00: there was a research published and then my team also validated it.
- 2:37SPEAKER_00: This was not our original research,
- 2:39SPEAKER_00: but we then went ahead and validated it and we saw a couple more things.
- 2:43This is around tool poisoning attacks against agents using MCP,
- 2:49model context protocol.
- 2:51SPEAKER_00: So you're able to actually do agent hijacking or fool the AI model into
- 2:57SPEAKER_00: exfiltrating data or stealing credentials as well.
- 2:59So you're going to see more and more of these newer attacks
- 3:04happening against agent AI piece.
- 3:07SPEAKER_00: But if I talk about the positives,
- 3:10SPEAKER_00: it does help SOC teams,
- 3:13SPEAKER_00: security teams,
- 3:14SPEAKER_00: IT teams become more efficient.
- 3:18SPEAKER_00: The agentic AI technology,
- 3:19SPEAKER_00: in my opinion,
- 3:20does reduce a little bit of risk of hallucinations and the
- 3:26false positives that
- 3:27SPEAKER_00: we're seeing with a single agent system,
- 3:30SPEAKER_00: right?
- 3:30SPEAKER_00: Maybe the last year was single agent system.
- 3:32SPEAKER_00: Now you have...
- 3:34agent doing something else,
- 3:35an agent doing validation of what this agent did.
- 3:38And then there is this maybe some automation that is bringing
- 3:41in context from the internet.
- 3:43SPEAKER_00: And then there is an agent that's putting all of that together.
- 3:46So the chance of issues,
- 3:49errors happening has definitely reduced,
- 3:53but it's still not foolproof.
- 3:54SPEAKER_00: So it should be leveraged.
- 3:56SPEAKER_00: It should not replace humans, not at this time.
- 3:58SPEAKER_02: Yes.
- 3:59And do you think there's a risk of businesses kind of rushing
- 4:01into it and missing
- 4:03SPEAKER_02: some security steps?
- 4:05SPEAKER_00: 100% that is already happening.
- 4:06And this is where it's a question is whether you know it or you
- 4:10don't know it,
- 4:11right?
- 4:12Because your employees are already using these different AI
- 4:15technologies and your
- 4:17SPEAKER_00: data may be going.
- 4:18SPEAKER_00: So it is very important for organizations to have that.
- 4:22SPEAKER_00: security and governance framework define?
- 4:24What are the policies around how they're going to secure the
- 4:27data that is being
- 4:28SPEAKER_00: sent to these AI applications?
- 4:30SPEAKER_00: What are those sanctioned AI applications for their environment?
- 4:34SPEAKER_00: What are they doing to,
- 4:35SPEAKER_00: if they're doing internal AI application development,
- 4:37SPEAKER_00: what are they doing to secure that environment as well?
- 4:40Because there are adversaries that will go after these
- 4:43applications to poison the outcomes.
- 4:46SPEAKER_00: So there's a plethora of different attacks I will see,
- 4:49but what that would mean is the security and governance
- 4:53framework should account
- 4:54SPEAKER_00: for that when building and investing in AI.
- 4:58SPEAKER_02: I want to talk as well about the phishing report.
- 5:00SPEAKER_02: So, you know, we talk every year about the phishing trends.
- 5:03SPEAKER_02: What have you seen this year?
- 5:04SPEAKER_02: What's the sort of highlights of the report that you put out recently?
- 5:06SPEAKER_00: Yeah, no, so fishing in terms of volume actually went down.
- 5:11SPEAKER_00: We saw about 20% reduction globally, but we're not surprised.
- 5:17We feel like there is definitely a shift happening from
- 5:20volumetric to quality,
- 5:22right?
- 5:22SPEAKER_00: Right.
- 5:24SPEAKER_00: more targeted,
- 5:25SPEAKER_00: more spear phishing attacks,
- 5:27more attacks involving bad guys bringing in the current context
- 5:32before they go
- 5:33SPEAKER_00: after an organization.
- 5:34SPEAKER_00: And because they're able to leverage automation,
- 5:36SPEAKER_00: they're able to leverage AI,
- 5:37SPEAKER_00: they're able to do it fairly accurately and at scale.
- 5:41SPEAKER_00: An example that I was just talking about with someone else was...
- 5:47SPEAKER_00: They will figure out when is your appraisal cycle.
- 5:50And then they will send out a CEO equity grant email to some of
- 5:54the employees that,
- 5:55SPEAKER_00: hey,
- 5:55SPEAKER_00: you got an equity grant as part of this appraisal cycle.
- 5:57SPEAKER_00: Now,
- 5:59SPEAKER_00: if that equity grant happens anytime during the year,
- 6:03SPEAKER_00: it may not result in employees clicking through.
- 6:05SPEAKER_00: But if it happens as part of the appraisal cycle,
- 6:07SPEAKER_00: suddenly you have that additional context and humans may fall for it.
- 6:11SPEAKER_00: Yeah.
- 6:12The other thing is we're going to see a lot of hybrid
- 6:15approaches where the
- 6:16SPEAKER_00: traditional means of sending phishing,
- 6:18SPEAKER_00: which is email,
- 6:19SPEAKER_00: SMS,
- 6:20SPEAKER_00: that will happen.
- 6:21SPEAKER_00: But in combination with voice-based phishing,
- 6:24SPEAKER_00: video-based phishing,
- 6:25SPEAKER_00: to,
- 6:26SPEAKER_00: again,
- 6:27SPEAKER_00: bring in that additional confidence on the end user.
- 6:30SPEAKER_00: Oh, okay, I saw this.
- 6:31SPEAKER_00: Oh, then I saw this as I learned that.
- 6:33SPEAKER_00: This looks believable.
- 6:34SPEAKER_00: Let me click on it or let me do this.
- 6:36And we're seeing examples of attacks becoming successful
- 6:40because they're doing
- 6:40SPEAKER_00: these hybrid approaches.
- 6:42SPEAKER_02: And are attackers using AI to help get more targeted,
- 6:47SPEAKER_02: or is it just more these trends we were seeing before AI was...
- 6:51SPEAKER_00: No,
- 6:51SPEAKER_00: they're definitely using AI.
- 6:53SPEAKER_00: Now,
- 6:53SPEAKER_00: it's hard to attribute,
- 6:55SPEAKER_00: but the reason I call this out is the email,
- 7:01SPEAKER_00: the quality of emails that are being sent out now are very,
- 7:05SPEAKER_00: very accurate.
- 7:06SPEAKER_00: They're able to do it in different languages.
- 7:09SPEAKER_00: Again, that's where I'm pretty sure they're leveraging AI.
- 7:11SPEAKER_00: Right.
- 7:12couple of the cases we actually took the text and we there is a
- 7:16ai detector app out
- 7:17there as well that will be able to tell you whether this text
- 7:19was generated using
- 7:20ai and and rightly so we found that yeah um so ai is being used
- 7:27to generate this ai
- 7:28is being used to generate voice space uh i think last year we
- 7:33showed you how
- 7:34SPEAKER_00: Jay's voice was being used.
- 7:36SPEAKER_00: I know Jay didn't make that call,
- 7:37but Jay's voice was used to generate a 10 seconds message to
- 7:43call an employee and
- 7:45SPEAKER_00: switch to text.
- 7:46SPEAKER_00: Yeah.
- 7:47SPEAKER_00: Video piece,
- 7:48we haven't seen much except for that news where some company
- 7:51got using Zoom-based
- 7:55SPEAKER_00: defake.
- 7:56SPEAKER_00: Yeah.
- 7:56SPEAKER_00: But hey, it can happen more and more.
- 7:58SPEAKER_02: Yeah.
- 7:59SPEAKER_02: Yeah.
- 8:01SPEAKER_02: It's pretty incredible, really.
- 8:02And there was also some news in the report about fishing
- 8:06hotspots,
- 8:07Brazil and Hong Kong.
- 8:08What are the sort of trends you're seeing there and what are
- 8:11the reasons behind
- 8:11SPEAKER_02: that,
- 8:12SPEAKER_02: do you think?
- 8:13SPEAKER_00: Look, fishing is a global issue.
- 8:16SPEAKER_00: I don't think certain regions are more susceptible to it than the other.
- 8:23The attackers will go after countries where there are these
- 8:26large companies,
- 8:26SPEAKER_00: which is where U.S.
- 8:27SPEAKER_00: still is number one.
- 8:30And then there are a lot of countries in the European region
- 8:32that you will see
- 8:33SPEAKER_00: getting targeted as well.
- 8:35And then there are up and coming economies where more and more
- 8:39manufacturing,
- 8:40SPEAKER_00: industrial stuff is happening.
- 8:42SPEAKER_00: So you will see them trying to get in over there.
- 8:45SPEAKER_02: chasing the opportunities.
- 8:46SPEAKER_00: Exactly.
- 8:47SPEAKER_00: Phishing is just a starting point.
- 8:48SPEAKER_00: A lot of the stuff that I just described,
- 8:51SPEAKER_00: once they're able to compromise an identity or an asset,
- 8:56SPEAKER_00: they will then attempt to move laterally.
- 8:57In many of the cases,
- 8:58they will sell that access to ransomware operators,
- 9:02deploy ransomware.
- 9:03The goal is eventually to gain more profits from each of these
- 9:07attacks that are performed.
- 9:09SPEAKER_02: And,
- 9:09SPEAKER_02: you know,
- 9:10SPEAKER_02: the,
- 9:10every year I think I've asked a question on advice for stopping
- 9:13fishing and it
- 9:13SPEAKER_02: comes back to zero trust.
- 9:15So it'd be great to talk about that and how businesses can kind
- 9:17of get ahead of these.
- 9:18SPEAKER_00: Yeah, no, I'll give you three things over here.
- 9:21SPEAKER_00: So look, and then we can make it broader than phishing, right?
- 9:25SPEAKER_00: AI-driven threat landscape that we're seeing.
- 9:27SPEAKER_00: There's tons of new stuff that we're going to see over next few months,
- 9:32SPEAKER_00: years as this,
- 9:34SPEAKER_00: because last year we weren't talking about a lot of agentic stuff.
- 9:37SPEAKER_00: This year is all about agentic stuff.
- 9:39SPEAKER_00: We're going to see more and more stuff getting developed, innovated.
- 9:43SPEAKER_00: Yeah.
- 9:45SPEAKER_00: Bad guys are already using AI to do a lot of the threat activity,
- 9:49SPEAKER_00: whether it's phishing,
- 9:49SPEAKER_00: malware generation,
- 9:50SPEAKER_00: exploitation,
- 9:51SPEAKER_00: recon activity,
- 9:53SPEAKER_00: post-infection activity as well.
- 9:56My advice from the defense perspective is you must invest in a
- 10:00platform that is
- 10:01SPEAKER_00: leveraging AI to fight AI.
- 10:03But we as an industry are not at a place where AI can replace
- 10:11humans or animals.
- 10:14AI cannot be that silver bullet that will solve the threat
- 10:17detection problem.
- 10:18SPEAKER_00: It is just going to make you more efficient.
- 10:20SPEAKER_00: It is going to improve the efficacy of the platform,
- 10:22SPEAKER_00: which is where you need to invest in a platform that uses AI to fight AI.
- 10:26SPEAKER_00: But the more important thing, as you mentioned, is...
- 10:30How do you shut down the attack paths and the vectors that
- 10:34these bad guys are using
- 10:36SPEAKER_00: or AI will use to attack the organization?
- 10:40SPEAKER_00: And you're able to do that by implementing zero trust architecture.
- 10:45You're able to do that by simplifying your network,
- 10:47remove complexities and
- 10:50essentially have a I mean the world that we imagine is there's
- 10:55no public IP
- 10:56addresses so the attackers can't attack what they can't see and
- 11:01your internal
- 11:03network is just like a guest Wi-Fi so fully segmented even if
- 11:08one of the identity
- 11:09or system is compromised you can't move around laterally which
- 11:12is a true definition
- 11:14SPEAKER_00: of Zero Trust Architecture
- 11:15SPEAKER_02: Yeah, absolutely.
- 11:18SPEAKER_02: And the move to agentic AI, it doesn't really change those fundamentals.
- 11:22SPEAKER_00: It absolutely doesn't, yes.
- 11:25SPEAKER_00: Look, those are all going to be applications now.
- 11:28SPEAKER_00: Leveraging those agentic AI applications to improve efficacy,
- 11:32SPEAKER_00: efficiency,
- 11:33SPEAKER_00: everyone will do it,
- 11:34SPEAKER_00: whether it's from the SOC side.
- 11:36We are developing our own apps to capitalize on the data that
- 11:41the Zero Trust Exchange produces.
- 11:43SPEAKER_00: We have 500 trillion daily signals,
- 11:45SPEAKER_00: 9 billion threats and policy violations on a daily basis,
- 11:49SPEAKER_00: half a trillion logs going through transactions,
- 11:52SPEAKER_00: going through our cloud.
- 11:53Ultimately,
- 11:54the AI solutions that would be more effective is where there is
- 11:56data.
- 11:58Your solution is as good as the data that you're going to
- 12:00leverage to train it.
- 12:01SPEAKER_00: So that's where we are actively investing in AI-driven operations.
- 12:07SPEAKER_00: So when you will visit Zenith Lab,
- 12:10SPEAKER_00: you will notice that inline zero trust everywhere,
- 12:13SPEAKER_00: that is number one priority still.
- 12:15We want to make sure we're able to help organizations implement
- 12:18zero trust everywhere.
- 12:20SPEAKER_00: Users, workloads, IoT, IoT, cloud.
- 12:23SPEAKER_00: But then leveraging all those insights,
- 12:26we're building these AI-driven operations where it will perform
- 12:30IT operations,
- 12:31SPEAKER_00: SOC operations,
- 12:33SPEAKER_00: leverage these data to train that,
- 12:35and then have policy decisions fed back into the platform as
- 12:39well in certain
- 12:40SPEAKER_00: scenarios.
- 12:41SPEAKER_00: So, yeah, that's the direction we're heading in.
- 12:43SPEAKER_02: Fantastic.
- 12:44SPEAKER_02: What are you most excited about in the security space at the moment?
- 12:48SPEAKER_00: Excited as well as worried, both at the same time.
- 12:51SPEAKER_00: I know you guys are hearing about AI a lot.
- 12:54SPEAKER_00: Look, there is a lot of potential in the agentic AI piece.
- 12:59I think it's going to take a good couple years to remove the
- 13:03noise and the fluff
- 13:05SPEAKER_00: and then start,
- 13:06SPEAKER_00: okay,
- 13:06SPEAKER_00: here are the three places where we are seeing real benefits.
- 13:10SPEAKER_00: Now,
- 13:10SPEAKER_00: you will see a lot of vendors say,
- 13:12SPEAKER_00: oh,
- 13:12SPEAKER_00: we're seeing benefits over here,
- 13:14SPEAKER_00: here,
- 13:14SPEAKER_00: here already,
- 13:15SPEAKER_00: but
- 13:16SPEAKER_00: I think we're still early on in many of these areas.
- 13:19SPEAKER_00: So next couple of years,
- 13:21I'm actually excited to see where it actually delivers on what
- 13:26promises we've been
- 13:27SPEAKER_00: seeing over the past few months.
- 13:29SPEAKER_02: Yeah, absolutely.
- 13:30SPEAKER_02: And how do you think that trend is going to go?
- 13:32SPEAKER_02: How do you sort of predict it will shake out?
- 13:36SPEAKER_00: I personally think it's going to keep going up.
- 13:39SPEAKER_00: The whole, it may not be agentic AI next year, it may be something else.
- 13:45But it's a cycle that we're seeing where there is rapid
- 13:49innovation happening.
- 13:51SPEAKER_00: And
- 13:53SPEAKER_00: The goal,
- 13:53SPEAKER_00: again,
- 13:53is how do we capitalize on the scale at which these LLMs are
- 14:01allowing us to operate
- 14:02SPEAKER_00: with,
- 14:02SPEAKER_00: right?
- 14:02SPEAKER_00: So there will be innovation on the compute side.
- 14:07SPEAKER_00: You're already seeing CPU being used instead of GPU, one bit.
- 14:11SPEAKER_00: I know that is still a little bit of marketing, but...
- 14:15SPEAKER_00: So you will see stuff on that.
- 14:17SPEAKER_00: You'll see innovation on training of these LLM sites as well,
- 14:21SPEAKER_00: which is what we saw with DeepSeek coming out.
- 14:25SPEAKER_00: Before that, what was the case?
- 14:27SPEAKER_00: There are these builders like...
- 14:29SPEAKER_00: OpenAI, Cloud, Llama, right?
- 14:33Then there are these adopters who will use those base models
- 14:38and then build on top
- 14:39SPEAKER_00: of it,
- 14:39SPEAKER_00: but they don't have ability to train the model from scratch.
- 14:43SPEAKER_00: That's these builders.
- 14:45And then there are application developers that will take what
- 14:48these adopters are
- 14:51SPEAKER_00: doing because they're providing an abstraction on top of it.
- 14:54SPEAKER_00: With DeepSeek,
- 14:54SPEAKER_00: what has happened is the pillar and the adoption folks,
- 14:58SPEAKER_00: the line has basically gone away.
- 15:01SPEAKER_00: You're able to now train with much less compute.
- 15:06SPEAKER_00: That research basically showed that.
- 15:07SPEAKER_00: So we will see innovation across the board,
- 15:11SPEAKER_00: hardware,
- 15:12SPEAKER_00: software,
- 15:13SPEAKER_00: and how do you build applications on top of it.
- 15:15SPEAKER_02: Fantastic.
- 15:16SPEAKER_02: Thank you so much, Stephen.
- 15:17SPEAKER_02: Really appreciate it.
- 15:18SPEAKER_00: Thank you.
- 15:18SPEAKER_00: Thank you.
- 15:21Bye.