[00:00:00] [music]
[00:00:04] >> Hi everybody. Today our guest Andre
[00:00:06] Andrew who is the CEO of Costella
[00:00:08] Intelligence and has been a CISO four
[00:00:11] times across 30 years in the field. He
[00:00:13] built Intercept systems for federal law
[00:00:15] enforcement and picked up three
[00:00:17] Department of Justice Awards while doing
[00:00:19] all of that. He has also been the chief
[00:00:21] application architect at Ogilvy and
[00:00:23] Mather and co-founded Bay Shore Networks
[00:00:25] as a CTO and took it through right
[00:00:27] through the acquisition parts. He’s also
[00:00:28] the author of the CISO Playbook and it’s
[00:00:31] follow up on the adversarial mindset.
[00:00:33] Holds multiple patents and outside of
[00:00:35] work he’s a international level
[00:00:37] certified judo coach. So that’s
[00:00:39] interesting. Andre, welcome to Tech on
[00:00:41] Inch. Great to have you on the show.
[00:00:42] >> Thanks for having me Ashar.
[00:00:43] >> All right. So all that security, I think
[00:00:46] you literally took it to a physical
[00:00:47] level of being a judo certified coach as
[00:00:50] well, right?
[00:00:51] >> It’s the reverse.
[00:00:52] >> Okay.
[00:00:53] >> I started judo in 1982, my first year in
[00:00:56] high school.
[00:00:57] >> Okay. So you kind of took the physical
[00:00:59] to the digital then.
[00:01:00] >> It’s funny because there are a lot of
[00:01:02] I’ll call them lessons. A lot of lessons
[00:01:04] in judo that transfer so well into the
[00:01:08] physical realm and life and business and
[00:01:11] it’s definitely been beneficial to me in
[00:01:13] my life, yeah.
[00:01:13] >> Okay. I’d love to hear a few examples or
[00:01:17] what kind of things kind of pass through
[00:01:18] before, you know, starting with the
[00:01:20] other questions.
[00:01:20] >> So in judo you learn how to take
[00:01:23] somebody’s energy, aggression, strength
[00:01:25] and use it against them. And so like for
[00:01:27] instance, I’m not a I’m not a big guy.
[00:01:29] I’m 5’4. And so when you work with a
[00:01:31] bigger person, they always instinctively
[00:01:33] they just try to muscle you. And when
[00:01:35] you reverse that on them and they fly,
[00:01:37] it’s kind of comical actually. But my
[00:01:39] point is in for example in business when
[00:01:42] you deal with someone very aggressive,
[00:01:44] if you go, you know, head to head like
[00:01:46] two bulls, you’re not going to get
[00:01:47] anywhere. So you need to learn how to I
[00:01:50] don’t want to say manipulate, but you
[00:01:51] know, really it is manipulation. How to
[00:01:53] manipulate where they want to take
[00:01:55] things and still end up with some
[00:01:56] benefit for yourself. And so, I think
[00:01:59] that in terms of dealing with that type
[00:02:01] of forward pressure and trying to push
[00:02:03] back all the time, learning how to get
[00:02:05] around it. That’s definitely been a less
[00:02:07] a physical lesson that translated into
[00:02:09] the business world.
[00:02:09] >> That’s interesting. Our topic generally
[00:02:11] today is around agentic AI or
[00:02:14] decentralized agentic AI in enterprise
[00:02:16] systems. And I mean, my first question
[00:02:18] really to start off is to really
[00:02:20] understand, are you currently deploying
[00:02:22] or seeing kind of AI agents being used
[00:02:25] effectively in production? And what kind
[00:02:27] of specific problems are they solving?
[00:02:29] >> Yeah, so I definitely agentic ecosystems
[00:02:32] exist in production. I will humorously
[00:02:34] state this next sentence, especially
[00:02:37] given the the name of your podcast
[00:02:39] Unhinged. So, I I do as [clears throat]
[00:02:41] I talk about this, I do want to separate
[00:02:43] what I could consider real agentic AI
[00:02:46] >> Yes.
[00:02:47] >> versus what marketing theater has
[00:02:49] labeled agents.
[00:02:50] >> And hyped it up, right? So.
[00:02:52] >> Oh, it’s it’s ridiculous. But, you know,
[00:02:55] on the security side especially, I see
[00:02:57] agentic ecosystems spawning up on the
[00:03:00] operational side all over the place. And
[00:03:02] it makes sense. You’re dealing with
[00:03:03] large volumes of data, distributed data
[00:03:05] sets, and there’s no human analyst that
[00:03:08] could go through all that volume. So,
[00:03:09] agents make sense in that type of an
[00:03:12] ecosystem. But, I will say this, where I
[00:03:14] see them most valuable are environments
[00:03:17] where bounds can be established clearly.
[00:03:20] I think that’s important, you know,
[00:03:21] where objectives are clear.
[00:03:23] >> Mhm.
[00:03:23] >> The agent can access data very, let’s
[00:03:26] say, cleanly, right? Because accessing
[00:03:28] data can become a little muddy. Missions
[00:03:30] can get controlled. And where the
[00:03:31] actions that an agent takes can be
[00:03:33] observed. And so, obviously, security
[00:03:35] ops, like I said, is a is a great place
[00:03:38] for that type of technology. But, I
[00:03:39] think you’ll see you’re going to see a
[00:03:41] lot of adoption of the technology in IT
[00:03:43] operations, not just security. You’re
[00:03:45] going to see them in security in
[00:03:47] software development, because this makes
[00:03:49] total sense in terms of, let’s say,
[00:03:51] augmenting and/or expanding software
[00:03:53] engineering operations. In other
[00:04:50] And that’s where I think, you know, I
[00:04:52] can tell you that at Constella, we’ve
[00:04:54] used agents to basically amplify our
[00:04:58] reach in terms of breach hunting because
[00:05:00] our breach hunting is at the core of the
[00:05:02] company. And so, for example, when I got
[00:05:04] to the company, this company was
[00:05:06] monitoring 400 Telegram channels, if you
[00:05:08] know what Telegram is, right? Once we
[00:05:10] went Ajanity, we now monitor 11,000
[00:05:13] channels. From 400 to 11,000 without
[00:05:16] adding any extra human. That’s the power
[00:05:18] of of Ajanity ecosystems in my opinion.
[00:05:21] >> Oh, that’s true. Absolutely. And you
[00:05:22] know, when we’ve been working with
[00:05:24] clients a lot of times, so one is being
[00:05:26] able to agent really is a somebody
[00:05:29] something or I don’t know to call thing
[00:05:31] or what do you call it, but an agent
[00:05:33] really and it reminds me of Matrix, by
[00:05:34] the way. So, I guess the word really
[00:05:36] came from there as well. Is autonomous
[00:05:38] decision-making, right? So, as long as
[00:05:39] your workflow or whatever you’re
[00:05:41] automating requires a certain level of
[00:05:44] autonomous decision-making, which is
[00:05:46] very difficult to write rules for or
[00:05:49] specific rules for, then yes, then you
[00:05:51] need it, right? So, I mean, when when we
[00:05:53] talk to clients, when they come up to
[00:05:55] and you know, they’re saying that, “Hey,
[00:05:57] I want to integrate AI agents or deploy
[00:05:59] those.” A lot of times, we just tell
[00:06:01] them that, “Hey, you don’t even need an
[00:06:02] agent, right?” So, this is like good old
[00:06:04] automation software. You can build a
[00:06:06] decision tree, you can build the logic,
[00:06:08] and just run it. So, you just pay once
[00:06:10] for developing that code base, rather
[00:06:12] than paying every time for an LLM cost
[00:06:14] as well, and everything and you know,
[00:06:16] you really don’t need an agent for it.
[00:06:17] >> You touched on something super
[00:06:18] interesting and super important from my
[00:06:20] perspective. You know, the autonomy
[00:06:22] within the decision-making process,
[00:06:24] because that leads to one of two
[00:06:26] pathways. Either the let’s just use the
[00:06:29] term agent loosely. Either the agent can
[00:06:31] make a decision and act on it, because
[00:06:34] we make decisions for action, right? It
[00:06:36] makes a decision and then goes into some
[00:06:38] type of waiting process, where a human
[00:06:41] or some other entity interjects and then
[00:06:43] takes that decision. Those are two very
[00:06:45] different pathways with very different
[00:06:47] levels of criticality attached to each
[00:06:49] one. And I think people are starting to
[00:06:52] realize that in theory, cuz it sounds
[00:06:55] great. My agent makes a decision and it
[00:06:57] takes action. I don’t have to do
[00:06:58] anything. That sounds great in theory.
[00:07:00] But when you put that into action,
[00:07:01] depending on what environments you play
[00:07:03] in, there are implications. So, I’ll
[00:07:05] give you an example. When I was at
[00:07:07] Bayshore Networks, we dealt with OT or
[00:07:09] ICS systems, right? Physical equipment.
[00:07:11] If you had any type of decision-making
[00:07:14] technology that did not have what’s
[00:07:16] called line of sight. In other words,
[00:07:18] that I can see where the action’s taking
[00:07:20] place, you put human lives at risk. So,
[00:07:22] for example, we dealt with a robotics
[00:07:25] plant that built cars. Those robots are
[00:07:27] if you’ve ever never seen them, they’re
[00:07:29] enormous. I mean, enormous. They’re like
[00:07:31] 12 high, you know, and they swing an arm
[00:07:33] across. If a human’s standing there,
[00:07:35] it’s a wrap. The outcome is good. And
[00:07:37] so, it was always interesting to me,
[00:07:38] because the administrators of certain
[00:07:41] robot robotics equipment were overseas
[00:07:44] and they would remotely come in to take
[00:07:45] actions on the robots. But they had no
[00:07:47] line of sight. You know, how how is that
[00:07:49] any different from an agent doing it?
[00:07:51] Does the agent actually have a sensor
[00:07:53] that sees that there’s a human inside of
[00:07:54] a cage where the robot is?
[00:07:56] >> Right.
[00:07:56] >> See what I’m saying? The implications
[00:07:57] start to get really interesting even
[00:08:00] though like I said in theory it sounds
[00:08:01] like, “Oh my god, this is amazing. It’s
[00:08:02] going to do all my work for me.”
[00:08:03] >> Correct. No, absolutely. That part is
[00:08:05] very important and very critical. Having
[00:08:07] AI constantly doing that thinking,
[00:08:09] running all of those tasks, that also
[00:08:11] can get get expensive really fast,
[00:08:13] right? So, what are your thoughts on
[00:08:15] that? What are the best ways to manage
[00:08:17] it? Have you tried something that
[00:08:18] worked?
[00:08:19] >> I think we’re moving, you know, at one
[00:08:20] point the digital transformation
[00:08:23] initiatives lived in a hybrid model and
[00:08:25] then everybody went cloud cuz we were
[00:08:27] going to save all kinds of money. Well,
[00:08:29] that’s turned out not to really be
[00:08:31] entirely true, right? And so, I think
[00:08:34] we’re going to end up back in a hybrid
[00:08:36] model. I think cloud technology is
[00:08:37] fantastic. I think it’s necessary given
[00:08:40] modern-day business, but I think there’s
[00:08:42] a hybrid nature that does introduce some
[00:08:45] security benefit and some cost benefit
[00:08:47] as well. And so, I think we’re going to
[00:08:49] end up in these hybrid spaces which also
[00:08:51] it has an implication in terms of what I
[00:08:53] like to label as intelligence hierarchy
[00:08:55] because for example, if you run frontier
[00:08:57] models which are very expensive. If you
[00:08:59] run a frontier model against every
[00:09:01] problem that you have, that makes I
[00:09:03] don’t know. That makes as much sense as
[00:09:05] a supercomputer to be a calculator,
[00:09:07] right? And so, you know, again it’s easy
[00:09:09] on paper to just go, “Okay, throw all of
[00:09:11] our problems at that ecosystem.” Not
[00:09:13] realizing it’s going to cost you a
[00:09:14] fortune to make simple decisions. And
[00:09:16] so, some problems just don’t require
[00:09:18] that type of capacity and so
[00:09:19] >> You don’t kill a You don’t kill a fly
[00:09:21] with a tank every time, right? So.
[00:09:23] >> That’s right. And and but it’s easy to
[00:09:24] make that mistake. I can see how any
[00:09:26] organization can just go, “Okay, yeah,
[00:09:28] just throw them all over there, right?”
[00:09:29] So, I think I think when enterprises
[00:09:32] start in intelligently routing work
[00:09:34] based on complexity, sensitivity,
[00:09:36] latency, cost, all those things, a small
[00:09:39] language model running locally may may
[00:09:41] make sense. It may classify information
[00:09:43] properly and inspect telemetry and all
[00:09:45] that stuff, and that’s okay. There’s
[00:09:47] nothing wrong with being in that hybrid
[00:09:49] model, and I think there are economic
[00:09:51] benefits to doing that cuz I mean, come
[00:09:53] on, the economics of just being totally
[00:09:55] cloud-based and totally, you know, being
[00:09:57] token-based within AI ecosystems, that
[00:10:00] that can get very ugly very quickly.
[00:10:02] >> No, that’s true. So, the one of the
[00:10:03] solution is shifting A is routing to the
[00:10:06] right model based on the problem that
[00:10:08] you’re solving, and the second is it
[00:10:10] shifting to smaller local hardwares to
[00:10:12] run smaller models. Would that be a
[00:10:14] potential solution as well?
[00:10:16] >> Absolutely. I think it comes down to
[00:10:17] putting the right amount of
[00:10:19] intelligence, let’s just use that term
[00:10:21] loosely, at the right point within an
[00:10:23] architecture, and that assumes the
[00:10:24] architecture is designed properly. I
[00:10:26] think it is possible. I think a lot of
[00:10:28] organizations have come to terms with
[00:10:30] the fact that there is no silver bullet
[00:10:32] here cuz the cloud really got sold to us
[00:10:34] as a silver bullet. It’s not. And so,
[00:10:35] you have to architect your stuff
[00:10:37] properly. Look at every organization
[00:10:38] that did the lift-and-shift approach.
[00:10:40] They took, you know, internally hosted
[00:10:42] applications, put them in the cloud, and
[00:10:44] it was just a disaster. And so, maybe
[00:10:46] not a functional disaster, but a cost
[00:10:48] disaster. And so, you know, that digital
[00:10:50] transformation initiative needs
[00:10:53] re-architecting to leverage the
[00:10:55] technology they’re going to run on
[00:10:56] natively. I think going back to a hybrid
[00:10:59] model is inevitable.
[00:11:00] >> How do you make agents secure, right?
[00:11:01] So, doesn’t leak private data, doesn’t
[00:11:04] create major security holes, and
[00:11:07] specifically when they are taking
[00:11:09] actions.
[00:11:09] >> I will humorously answer your question
[00:11:11] by saying, you can’t make them secure
[00:11:13] enough. It’s just difficult to make them
[00:11:15] secure.
[00:11:16] >> Okay. So, you make your systems more
[00:11:18] secure, right? So, A or B.
[00:11:20] >> Yeah, yeah, you have to you you you have
[00:11:21] to take a holistic approach. You have to
[00:11:23] really look at the entire architecture
[00:11:25] to introduce your security or protection
[00:11:27] into this into this type of initiative.
[00:11:30] I think what’s interesting is if you
[00:11:32] think of it from a parallel perspective
[00:11:34] where you treat protecting agents like
[00:11:37] protecting an identity. And so, for
[00:11:39] example, an identity can authenticate
[00:11:41] into something, it can access data, it
[00:11:43] can invoke tools, etc. So can an agent,
[00:11:46] right? So can a piece of AI because not
[00:11:48] all AI is identic, but my point is from
[00:11:51] a security perspective, if you treat it
[00:11:53] as a privileged actor, that means you
[00:11:55] can start asking the same questions that
[00:11:57] you would of a human, right? Who is this
[00:12:00] agent? Who authorized this agent to be
[00:12:02] here? What can it access? What actions
[00:12:04] can it take? What data can it retrieve?
[00:12:06] And the most important question, how can
[00:12:08] I stop this thing? Because, for example,
[00:12:10] an identity, a human, if I realize that
[00:12:12] Ashar logged in from, let’s say,
[00:12:15] California into my system, but at the
[00:12:17] same time Ashar logged in from India, we
[00:12:20] have a problem. I need to kill those
[00:12:22] sessions cuz I don’t know which one is
[00:12:23] real.
[00:12:23] >> Which one is real? Yeah.
[00:12:24] >> Right? Unless you’re flipping VPNs
[00:12:26] globally and stuff, which is possible,
[00:12:28] but most humans don’t do that. Most
[00:12:30] normal people don’t do that. And so,
[00:12:31] killing those sessions is something that
[00:12:33] a security team should be or an identity
[00:12:35] team should be able to do at the click
[00:12:36] of a button. How do you stop an agent at
[00:12:39] the click of a button? And especially if
[00:12:40] you introduce a distributed identic
[00:12:43] ecosystem, now it’s not that easy. And
[00:12:45] so, I think agents should really be
[00:12:48] approached and it’s kind of like the
[00:12:49] zero trust notion, right? With a
[00:12:51] narrowly scoped, very short-lived set of
[00:12:53] permissions, right? Inheriting a giant
[00:12:56] set of permissions to access all the
[00:12:57] data in your ecosystem and stuff like
[00:12:59] that, that’s probably not a wise move.
[00:13:02] But, you know, I think ultimately we
[00:13:04] need we, as the defenders, need to
[00:13:06] create friction in those environments
[00:13:08] where we can step in, like I said, and
[00:13:10] kill something based on data that we’re
[00:13:12] receiving, telemetry that we’re
[00:13:14] receiving. And so, I think that
[00:13:16] ultimately that’s what this comes down
[00:13:17] to. But, securing this stuff is far from
[00:13:19] trivial. Let me sum it up that way.
[00:13:21] >> And it’s become more, like you said,
[00:13:23] it’s become more difficult, right? So,
[00:13:24] previously you had actions, you could
[00:13:26] control actions, is a you know, person
[00:13:28] clicking or something or some sort of
[00:13:30] attack coming in. With agent, you just
[00:13:31] don’t know. You give him in a right
[00:13:33] access, but
[00:13:34] >> So, imagine this now, because this is
[00:13:35] where I think it gets really, really
[00:13:37] fascinating. And this is just real-world
[00:13:39] perspective, right? Everything that we
[00:13:41] just talked about, great conversation.
[00:13:43] But, let’s be honest, the majority of
[00:13:45] enterprises, when they deploy agentic
[00:13:47] anything, they didn’t build that stuff
[00:13:49] in-house. They either brought
[00:13:50] consultants in or bought products that
[00:13:52] claim to do this. Do you really
[00:13:54] understand tech at a at a low-level
[00:13:56] tech, right? Do you really understand
[00:13:58] what that ecosystem is doing?
[00:13:59] >> No.
[00:13:59] >> You’re just you’re you’re taking stuff
[00:14:01] you’re taking a lot of things at face
[00:14:03] value because you have to trust your
[00:14:05] consultants and/or vendors. And, you
[00:14:06] know, there’s a reality to that. So, how
[00:14:08] do you even secure something that you
[00:14:09] don’t truly understand?
[00:14:11] >> No, that’s true. And that’s why you rely
[00:14:13] on compliance systems, right? So, you
[00:14:14] would ask them, “Are you ISO compliant?
[00:14:17] Are you SOC 2 compliant? Not compliant?”
[00:14:19] I mean, I I get what you’re saying, but
[00:14:21] and then that’s the way of enterprises
[00:14:22] justifying, “Hey, we checked that.”
[00:14:24] >> Yes, which and I’ve spoken about this,
[00:14:26] which I think is the grandest mistake
[00:14:29] this this entire industry has made is
[00:14:31] rely on compliance and think that that
[00:14:33] makes them secure.
[00:14:34] >> Right. So, how do they then What’s the
[00:14:36] alternative?
[00:14:37] >> The alternative is having properly
[00:14:38] skilled people on staff that can
[00:14:41] scrutinize technology and make decisions
[00:14:43] accordingly. But, what happens? That
[00:14:45] means that the deployment of of
[00:14:46] technology X goes from, let’s say, a
[00:14:49] 1-month deployment to, let’s say, 3
[00:14:51] months. Is the organization comfortable
[00:14:53] with that? Because this takes time. If
[00:14:55] you didn’t build the technology, you
[00:14:56] have to tear it apart a little bit to
[00:14:58] understand what it’s actually doing
[00:15:00] within your environment. How many
[00:15:01] organizations do that? I haven’t run
[00:15:03] across that many.
[00:15:04] >> Yeah, that’s true. So, what are the What
[00:15:05] have been some of the biggest security
[00:15:07] oversights that you’ve seen SAS, you
[00:15:09] know, by SAS organizations or outsource
[00:15:11] {slash} founders companies?
[00:15:13] >> I think the biggest mistake is that And
[00:15:15] this is And this again, this goes back
[00:15:17] 20-plus years in this industry, and I’ve
[00:15:19] always laughed at this. Security A lot
[00:15:22] of leaders think that because they put a
[00:15:24] policy in place that that means Think
[00:15:26] about it like this, I always say. Your
[00:15:27] policy is there to make your good people
[00:15:30] continue to be good. You think your
[00:15:31] attackers care about your policy? They
[00:15:33] don’t care what the policy says. That’s
[00:15:35] one. Two, how do you enforce that
[00:15:37] policy? Because most policies are what?
[00:15:39] A document that lives somewhere that you
[00:15:41] expect people to read and people to
[00:15:43] adhere to. I don’t know, last time I
[00:15:44] checked we have laws against crime and
[00:15:46] crime takes place all the time. That
[00:15:48] should tell you something about human
[00:15:49] nature.
[00:15:50] >> So, that’s true.
[00:15:52] >> Well, the model must never do X. That is
[00:15:55] That’s not security. That is a hope. And
[00:15:58] so, you know, I think if you call Now,
[00:16:00] let let me say Let me say this. If you
[00:16:02] compound that with giving, for example,
[00:16:04] a genetic ecosystems excessive
[00:16:06] privileges because it happens, because
[00:16:08] you want it to do all this fancy stuff
[00:16:10] and you need it to touch all these parts
[00:16:12] of your ecosystem. So, you give it
[00:16:13] excessive privileges and then all of a
[00:16:15] sudden you end up with these
[00:16:16] architectures that you assume the model
[00:16:19] should never do X. Yet, you gave the
[00:16:21] architecture the elements in the
[00:16:23] architecture all kinds of access and
[00:16:25] then we get surprised that we have
[00:16:26] security problem. I think, like I said,
[00:16:28] it comes down to really understanding
[00:16:31] what it is that’s being deployed and
[00:16:32] then making the proper decisions based
[00:16:34] on that understanding and my perspective
[00:16:37] also based on the understanding of who
[00:16:39] could potentially want to attack me
[00:16:41] through this. Then you can start having
[00:16:43] even some proper notions about
[00:16:45] protecting environment. Cuz until then,
[00:16:47] I’m sorry, it’s kind of a hope.
[00:16:48] >> Right. So, most of these Then what what
[00:16:50] we’re getting to is that, hey, most of
[00:16:52] these AI governance principles or
[00:16:53] security principles generally rely on
[00:16:55] compliance checklists and some sort of,
[00:16:57] you know, policy guardrails. You feel
[00:16:59] it’s not really effective. So, then
[00:17:01] what’s a way to re-engineer I know one
[00:17:05] thing is that we have in-house people,
[00:17:07] they start pen testing the entire
[00:17:08] software and going through, but that’s
[00:17:10] that’s a time that’s not, you know,
[00:17:12] feasible, right? So, if let’s say you’re
[00:17:13] deploying Salesforce, would you be going
[00:17:16] through all of that stuff?
[00:17:18] >> And that’s And that’s where it gets
[00:17:20] really challenging. I mean, look, you
[00:17:21] want to do this properly. First of all,
[00:17:23] you know, pen testing is great, but it
[00:17:25] exists within confines, within bounds,
[00:17:28] right? So, I would opt more for a true
[00:17:30] red team that can simulate This is
[00:17:33] important. That can simulate a
[00:17:35] particular adversary. Not all
[00:17:37] adversaries are going to be interested
[00:17:38] in your company. So, you got to figure
[00:17:40] out which ones are going to be
[00:17:41] interested and then simulate how they
[00:17:44] operate. Not that you’re telling, “Hey,
[00:17:47] don’t touch this.” Simulate it with
[00:17:49] people that will be just like the
[00:17:50] attacker.
[00:17:51] >> Doing the opposite. Right.
[00:17:52] >> Then, right, then you can start going,
[00:17:54] “Okay, wait a minute. We have a weakness
[00:17:56] here.”
[00:17:56] >> Right.
[00:17:57] >> My parallel is always this. If I’m going
[00:17:59] to build a security system for my home,
[00:18:01] a physical alarm system, I don’t want
[00:18:03] the alarm company to give me advisory
[00:18:05] and I don’t want a police officer cuz
[00:18:07] none of them think like the burglar.
[00:18:09] Bring me a burglar who will look at my
[00:18:11] house and go, “Yeah, I’d get in this
[00:18:13] way.” I go, “Okay.”
[00:18:15] >> [laughter]
[00:18:16] >> And so, you know, in in the digital
[00:18:18] space, we need a lot of the same. We
[00:18:20] need people that are going to come in
[00:18:22] They don’t care about our rules. They
[00:18:23] don’t care about our policies. They
[00:18:25] don’t care about our boundaries because
[00:18:26] the attacker does it. You know, I’m not
[00:18:28] saying this is easy, by the way. I I I
[00:18:30] mean, hell, I know for a fact, I’ve
[00:18:32] lived it. It’s not easy at all, but it
[00:18:34] is the more responsible way of doing
[00:18:36] things.
[00:18:37] >> Now, you as a leader need to decide, “Do
[00:18:39] I want the easy route? You know, do I
[00:18:40] want the route where there’s no
[00:18:42] friction?” Because that’s what they come
[00:18:43] in, right? A lot of leaders
[00:18:45] >> come into an environment and they go,
[00:18:46] “Oh, we got our SOC 2 report.” And by
[00:18:48] the way, SOC 2 is not compliance, so
[00:18:50] let’s just put that on the table. You’re
[00:18:52] not complying to anything. You have a
[00:18:53] SOC 2 report or you have your ISO 27001
[00:18:56] certificate and then you present that
[00:18:58] upstream to leadership and maybe the
[00:18:59] board cuz a lot of security leaders
[00:19:01] don’t report to the board. That is a
[00:19:03] myth. And they go, “Okay, you know,
[00:19:04] since we’re covered here, we’re good.”
[00:19:06] Well, I bring you to the example of
[00:19:07] Uber. Uber has every certification badge
[00:19:10] on the planet and it’s been breached how
[00:19:12] many times?
[00:19:13] >> Several times.
[00:19:13] >> Yeah, so that tells you you can have all
[00:19:15] of them. It still doesn’t mean that
[00:19:17] you’re secure. You know, look, let’s be
[00:19:18] real. Savvy attackers when they build a
[00:19:20] campaign against your environment, there
[00:19:22] is generally generally, not always, but
[00:19:25] there’s generally a human component. In
[00:19:26] other words, they’re targeting humans in
[00:19:28] your environment. How do you certify
[00:19:30] against human weakness? Like how do I
[00:19:32] start it? Let’s say you have an
[00:19:33] executive assistant, Ashar, and that
[00:19:35] person knows your entire life. If I want
[00:19:37] to attack you, I’m not going to come at
[00:19:39] you. I want your assistant.
[00:19:41] >> Right.
[00:19:41] >> Because your assistant is likely an
[00:19:43] easier pathway into me making you an
[00:19:45] attackable profile and then attacking
[00:19:48] you. So, I find their emotional
[00:19:50] weaknesses, I find, you know, their
[00:19:52] likes, dislikes, whatever. This is
[00:19:54] common tradecraft. I target that human.
[00:19:56] Like what certification could you put
[00:19:58] the company through to say, “Well, my
[00:20:00] executive assistant would never be
[00:20:01] breached.” That’s security reality and
[00:20:03] granted, I realize we were talking about
[00:20:05] agents and AI and this has nothing to do
[00:20:07] with it, but that is security
[00:20:09] >> reality.
[00:20:09] >> That’s true.
[00:20:09] >> I guess that’s that’s what, you know,
[00:20:11] you mean by when you write about
[00:20:13] adversarial mindset and thinking from
[00:20:14] that strategy. So, kind of, you know,
[00:20:16] that’s the concept.
[00:20:17] >> Yeah, I’m telling you as an adversary
[00:20:19] because okay, one of the reasons a lot
[00:20:21] of security leaders can’t put themselves
[00:20:24] in that mindset because they’ve never
[00:20:25] actually done security work. That’s one
[00:20:27] of the challenges of the modern day
[00:20:29] security leadership. They’ve actually
[00:20:30] never been offensive, they’ve never
[00:20:32] actually lived in the trenches.
[00:20:33] >> So, a really great hacker could actually
[00:20:36] be the best security consultant you can
[00:20:38] have on the board.
[00:20:40] >> Yeah. Yes, [laughter] you want to bring
[00:20:42] those people on your team so that they
[00:20:45] can they can tell you, “Dude, no.” I’ll
[00:20:48] give you an example. I worked at an
[00:20:49] organization and when I walked in, just
[00:20:52] to show you the mindset, when I walked
[00:20:53] in, I said, “All right, how did we spend
[00:20:55] our security budget?” Okay, we did the
[00:20:57] best EDR software ever. It’s on every
[00:21:00] laptop. Cool. We have network, you know,
[00:21:02] detection and response. Cool. I said,
[00:21:04] “What matters to this company?” When we
[00:21:06] started answering those questions, we
[00:21:08] realized that we had multiple
[00:21:09] customer-facing databases that were
[00:21:11] fronted by APIs and applications, right?
[00:21:13] And when I looked at how much of the
[00:21:14] security budget was spent protecting
[00:21:16] those things, it was virtually zero. But
[00:21:19] we had the most secure laptops on the
[00:21:20] planet. And then my next question
[00:21:22] obviously is, well, which one of those
[00:21:23] laptops has an entire copy of these
[00:21:25] massive databases? None. Okay, so you
[00:21:28] spent all your money protecting
[00:21:29] something that does not have the data
[00:21:31] that we want to protect. That That makes
[00:21:33] so much sense.
[00:21:34] >> Right?
[00:21:34] >> And so it’s not a judgmental statement.
[00:21:37] What it What it is is a statement
[00:21:38] establishing the fact that that leader
[00:21:40] did what made sense to him.
[00:21:42] >> Yeah.
[00:21:42] >> It was Alan in that case. It made sense
[00:21:44] to him. He was familiar with the
[00:21:46] endpoint world. He was familiar with
[00:21:48] what could happen through an endpoint.
[00:21:50] Nothing wrong with that. But they never
[00:21:51] took a look at this other side because
[00:21:53] it didn’t come natural to them. They
[00:21:54] didn’t come from the application world
[00:21:56] or the API world or the database world.
[00:21:58] All of a sudden you have these databases
[00:22:00] with all this critical data totally
[00:22:02] operating unprotected. You You see what
[00:22:04] I’m saying? And that’s why when I say
[00:22:06] cuz they had SOC 2 reports, they had SOC
[00:22:09] reports, they had all of those All of
[00:22:11] them were covered. Because what’s the
[00:22:12] first thing you could do when you get a
[00:22:14] compliance initiative? You can establish
[00:22:16] the the zone, the boundary. So what What
[00:22:18] happens if I go, well, we’re going to
[00:22:19] get a SOC 2 report on our endpoint
[00:22:21] ecosystem. Don’t touch the databases. We
[00:22:23] look great. We got a SOC 2 report. It
[00:22:25] says we’re awesome. Are we? You see
[00:22:26] where the problem starts to
[00:22:27] >> Yeah, yeah, absolutely. Absolutely.
[00:22:29] >> And so
[00:22:30] you know, to me you have to have people
[00:22:32] in the seats that can look across the
[00:22:34] board and go, wait a minute, that
[00:22:35] doesn’t make sense. We need to protect
[00:22:37] this too. And we need to protect it
[00:22:38] natively. Because okay, I’ve also worked
[00:22:40] in environments where the actual web
[00:22:42] applications had tight web app firewalls
[00:22:45] in front of them and they were totally
[00:22:46] protected. But the APIs that touched the
[00:22:48] same database were totally wide open.
[00:22:50] >> It’s really about understanding and you
[00:22:52] know, looking and writing implementing
[00:22:54] it at the right place. A lot of times
[00:22:56] the budget just goes in the wrong
[00:22:57] directions and
[00:22:58] >> I mean, one of the things you talked
[00:22:59] about has been using AI to spin up, you
[00:23:02] know, dynamic digital twins and smart
[00:23:04] honey pots to confuse attackers. How do
[00:23:07] security teams kind of set up those
[00:23:08] environments to help, you know, or act
[00:23:11] as a early warning tool?
[00:23:12] >> First of all, I’m a huge fan of
[00:23:13] deception, what I call deception
[00:23:15] engineering. You know, look, it allows
[00:23:16] you to control the terrain in such a
[00:23:18] fashion where, at the very least, you
[00:23:21] know what the adversary is seeing as
[00:23:23] opposed to them just having free reign.
[00:23:25] So, it comes down to one, and this goes
[00:23:27] back to one of the points I was just
[00:23:28] making. You have to understand your
[00:23:30] ecosystem an actual attacker would be
[00:23:32] interested in so that you can build
[00:23:34] mirrored environments of those
[00:23:36] ecosystems and make them functional
[00:23:38] enough that the attackers will think
[00:23:40] they’re actually touching real system.
[00:23:42] >> Right.
[00:23:42] >> And so, it still comes down to knowing
[00:23:44] your environment, and then yes, you can
[00:23:46] use tools to build digital twins and
[00:23:48] honey pots. Honey pots can be fantastic
[00:23:50] because they may not catch, per se,
[00:23:52] right? Quote unquote catch an attacker,
[00:23:54] but they certainly do create early
[00:23:56] warning systems where you know that
[00:23:58] nobody should have touched environment
[00:24:00] X, and if it got touched, somebody’s
[00:24:02] messing with your system, right?
[00:24:04] >> Right. Right.
[00:24:05] >> So, I think from an early warning
[00:24:06] perspective, honeynets, honeypots,
[00:24:08] honeynets, honey tokens, fantastic
[00:24:10] systems. From an intelligence gathering
[00:24:12] perspective to build your threat intel
[00:24:15] base, if you will, as as an
[00:24:16] organization, again, another tremendous
[00:24:19] source of information. What When again,
[00:24:21] when I was at Bayshore Networks, we did
[00:24:24] a number of operational technology OT
[00:24:27] honeypots, and we got fantastic value
[00:24:29] out of them. The Now, those operated off
[00:24:31] industrial ICS protocols, not regular
[00:24:34] like HTTP or FTP type protocols. Once
[00:24:36] you understood the protocols, you
[00:24:38] realized you could build systems that
[00:24:39] were very realistic. They seemed totally
[00:24:42] real, but they were totally synthetic.
[00:24:44] And so, we got a lot of value in real
[00:24:46] environments when we deployed them out
[00:24:48] there, you know, external entities, I’ll
[00:24:49] just call them, started touching them.
[00:24:51] We would get very great intel out of
[00:24:53] that. That informed the policies that we
[00:24:55] would then put into our devices to
[00:24:57] enforce on those network flows. From
[00:24:59] that perspective, that loop of
[00:25:01] intelligence to active protection, I
[00:25:03] think deception engineering is is
[00:25:05] unbeatable.
[00:25:06] >> No, absolutely it helps, right? So, you
[00:25:08] get to see
[00:25:09] the whole system and it’s a digital
[00:25:11] twin, you get to see how somebody’s
[00:25:13] approaching it, not approaching it, when
[00:25:15] it’s being in a you know, approached
[00:25:17] incorrectly, that kind of just makes it
[00:25:19] easier, just makes the overall job
[00:25:21] easier.
[00:25:21] >> Yeah.
[00:25:21] >> All right. Let’s talk a little about
[00:25:23] because you’ve kind of mentioned it in
[00:25:25] your post and a little bit online as
[00:25:28] well where what’s a problem a lot of
[00:25:30] time organizations inherit these legacy
[00:25:32] systems or have existing legacy old
[00:25:36] systems. So, what is your framework to
[00:25:38] work around with them, how to make them
[00:25:41] easier to manage, how to make them more
[00:25:43] secure, is there a certain framework
[00:25:44] that you kind of deploy or how do you go
[00:25:47] about that?
[00:25:47] >> I think you might talk to people that
[00:25:49] would answer that by saying I follow
[00:25:50] this framework. I think that’s absolute
[00:25:52] You have to understand your
[00:25:54] environment and if you don’t, there’s no
[00:25:57] way you could force fit a framework to
[00:25:59] your environment. I’ll give you an
[00:26:00] example. Let’s say I went into I got
[00:26:02] brought into a bank and this bank is
[00:26:04] still using a mainframe because banks
[00:26:07] still use mainframes, believe it or not.
[00:26:08] And let’s say it’s a VAX VMS system.
[00:26:11] Most people today don’t even know what
[00:26:12] that is, right?
[00:26:14] protections to a VAX VMS system? Well,
[00:26:17] if you don’t know VAX VMS, good luck.
[00:26:21] Right? Because there’s not a framework
[00:26:23] on the planet right now and I realize
[00:26:24] people love frameworks because they want
[00:26:26] everything to be wrapped up into a tight
[00:26:27] box with a bow on it. That’s not the way
[00:26:29] the real world works. And so, if I don’t
[00:26:31] come in with the right skill or the
[00:26:33] right team that can look at that
[00:26:35] ecosystem and say, oh, well, here are
[00:26:38] our points of risk, here are our attack
[00:26:40] pathways, our attack surface looks like
[00:26:42] this. Until I have answers to all those
[00:26:44] questions which are subjective to that
[00:26:46] environment, I can’t build a protective
[00:26:48] environment. Like what am I going to do?
[00:26:50] Go to any modern day security company
[00:26:52] and say, “Hey, protect my VAX VM.” When
[00:26:54] they go, yeah, half of the people that
[00:26:56] work there are going to go, “What is
[00:26:58] VAX?” Now, I’m telling you this is a
[00:26:59] real example, by the way, because I know
[00:27:01] that there’s a very large international
[00:27:04] bank that still runs everything on a VAX
[00:27:06] VMS back end. I know this for a fact.
[00:27:08] So, how do you protect that? And so, you
[00:27:10] really have to understand that ecosystem
[00:27:12] and look at all the touch points.
[00:27:14] They’re just And like I said, I know
[00:27:15] everybody wants to, you know, cleanly
[00:27:16] package it. It doesn’t work that way.
[00:27:18] You have to go in, you have to do your
[00:27:20] homework, do the analysis, and then
[00:27:22] build a protective set of mechanisms
[00:27:24] subjective to that environment. And I
[00:27:26] realize there are a lot of people that
[00:27:28] will listen to that answer and totally
[00:27:30] hate that answer. Well, give me one
[00:27:31] that’s not any more that’s any more real
[00:27:33] than what I just said.
[00:27:34] >> That’s exactly correct. Every legacy
[00:27:35] system would have its own set of
[00:27:37] problems. It would have I mean, you
[00:27:38] might be look You might have a framework
[00:27:40] to look at similar set of things, but
[00:27:42] the way you look at it, the way you fix
[00:27:44] it, the way you diagnose this, that all
[00:27:47] of that becomes different, right? So,
[00:27:48] you might be looking for a common set
[00:27:50] of, you know, potential security flaws
[00:27:53] or something that, hey, I need to check
[00:27:55] this and that and whatnot. So, passwords
[00:27:57] are not stored in the database, they’re
[00:27:58] encrypted, they are etc. etc. But, the
[00:28:01] way the solution turns out to be or the
[00:28:04] way you would improve it or make it more
[00:28:06] secure would always be kind of different
[00:28:08] or will always have some level of
[00:28:09] subjectivity.
[00:28:10] >> Again, real example from the OT from my
[00:28:13] work at the OT environments back uh some
[00:28:15] years ago. This is What I’m about to
[00:28:17] tell you is one of the reasons why
[00:28:19] people like me become very unpopular
[00:28:21] very quickly because we just tell you
[00:28:23] the the the unfortunate truth. So, I
[00:28:25] went into this environment and they had
[00:28:27] spent a ton. They brought in consultants
[00:28:30] to secure these OT environments, right?
[00:28:32] The effects motors, PLCs, all that type
[00:28:34] of equipment. Equipment that has
[00:28:36] physical impact on the world, right?
[00:28:37] >> Right.
[00:28:38] >> And what did these people do? They came
[00:28:39] in and they basically brought in IT
[00:28:41] technology. So, for example, they said,
[00:28:44] “Okay, this PLC can only be spoken to
[00:28:47] from this source IP address. Okay,
[00:28:49] that’s great. But you’re telling me that
[00:28:51] if I’m on the network as an attacker, I
[00:28:52] can’t spoof that source address, which
[00:28:55] by the way, I did. And they realized
[00:28:57] they spent millions of dollars building
[00:28:59] a protective mechanism that was
[00:29:00] virtually useless against a
[00:29:02] sophisticated attacker. My thesis to
[00:29:04] them was, what we should do is protect
[00:29:06] this at a protocol level where we can
[00:29:09] say, okay, this motor, for example,
[00:29:11] should never run past N RPMs. But you
[00:29:15] can’t do that if you’re just doing
[00:29:16] network segmentation. Network
[00:29:18] segmentation only says this these IP
[00:29:21] addresses can talk to these IP
[00:29:22] addresses. It doesn’t say if one of
[00:29:24] these actually sends a signal and tells
[00:29:26] the motor to operate outside of its
[00:29:28] operational constraints, stop that. You
[00:29:30] can only do that at the protocol level
[00:29:32] or at the ladder logic level at the end
[00:29:34] point, one of the two. My thing was,
[00:29:36] we’re not touching ladder logic cuz
[00:29:38] that’s dangerous, but I can on the
[00:29:40] network flow enforce set point range
[00:29:43] restrictions. So, what happened? That
[00:29:45] made me very unpopular because I was
[00:29:47] talking a language that most security
[00:29:48] people have no idea about. They were
[00:29:50] like, what are you talking about? That’s
[00:29:51] the right way to protect these
[00:29:53] environments. Now, it’s difficult, which
[00:29:55] usually equates, right? And the right
[00:29:57] way is usually harder to do, right? And
[00:29:59] so, a lot of organizations chose not to
[00:30:02] adopt that model because it was
[00:30:04] difficult, even though it was the better
[00:30:06] secure model. If you study an attack
[00:30:08] like Stuxnet, what happened in Stuxnet?
[00:30:10] The signals went across the network and
[00:30:12] told the centrifuge controllers to spin
[00:30:15] out of control and they exploded. That’s
[00:30:16] exactly what I was talking about. Nobody
[00:30:19] wanted to do it because it was very
[00:30:20] difficult. That’s the right way to
[00:30:22] secure those type of environments. And
[00:30:24] I’m just giving you this as a parallel
[00:30:25] to what we were just talking about
[00:30:27] because my point is, if I didn’t
[00:30:29] understand that environment at that
[00:30:31] granular level, I could never come up
[00:30:33] with a security strategy that would
[00:30:35] operate at a protocol level. I would do
[00:30:37] what those guys did. They’d bring an IT
[00:30:39] security solution into an OT
[00:30:41] environment. It doesn’t really work.
[00:30:43] Some benefit, don’t get me wrong, but
[00:30:44] it’s not a true end-to-end solution.
[00:30:47] >> So, how can, you know, CEOs, CTOs, CISOs
[00:30:50] kind of stop treating A lot of
[00:30:52] organizations treat security as some
[00:30:53] level of bottleneck, right? So, you
[00:30:55] know, my release is delayed because
[00:30:57] security still needs to be tested out.
[00:30:59] We want this out as quickly as possible,
[00:31:01] but security is stopping us or whatever.
[00:31:03] How What would you advise or what do you
[00:31:05] do say or structure it so that it
[00:31:07] instead of becoming a bottleneck,
[00:31:09] somehow these two teams align so that
[00:31:11] security eventually is helping product
[00:31:14] delivery speed up, right?
[00:31:15] >> Look, that comes down to the leader at
[00:31:17] hand. The leader has to tactfully,
[00:31:20] that’s important, tactfully inject
[00:31:22] themselves into all of the processes
[00:31:24] that are going to end up in with a
[00:31:26] release of anything. So, security has to
[00:31:28] be part of the natural process, right?
[00:31:30] If it’s an SDLC, security has to be part
[00:31:32] of that SDLC. If you can’t achieve that,
[00:31:34] we’ll never get past the point where we
[00:31:36] are now, which is basically either
[00:31:37] security is an afterthought, in other
[00:31:39] words, something got released and then,
[00:31:41] oh, by the way, we have security
[00:31:43] problems, or they come to security the
[00:31:45] day before the release, and you know, if
[00:31:47] you bring up any objections, you’re
[00:31:49] going to cause a business impact, which
[00:31:51] has a revenue impact, which guess what?
[00:31:53] You are going to lose because business
[00:31:55] and revenue will win all the time. And
[00:31:57] that’s okay, that’s natural, right? So,
[00:31:59] you as a leader have to realize that and
[00:32:01] have to make sure that you can get your
[00:32:04] teams injected into the life cycle of
[00:32:07] whatever’s being built so that security
[00:32:08] is inherently factored in as much as
[00:32:11] possible. I know it’s not going to be
[00:32:12] perfect, built into those to those
[00:32:14] iterations that lead to a release. That
[00:32:16] is the only way to start properly
[00:32:18] solving these problems.
[00:32:20] >> What I just said is extremely difficult
[00:32:21] to do, which is why a lot of security
[00:32:23] leaders come in and they go, well, am I
[00:32:25] going to go bang my head against that
[00:32:26] for 5 months or get a SOC 2 report in 2
[00:32:29] weeks and look like a rockstar? Well,
[00:32:32] guess what? This normally wins because
[00:32:34] they know that this is going to be way
[00:32:35] more difficult, but this is the right
[00:32:37] thing to do.
[00:32:37] >> Right.
[00:32:38] >> So, final question. What advice you
[00:32:40] would give to C-suite leaders today to
[00:32:42] ensure that the security culture kind
[00:32:44] of, you know, doesn’t get left behind
[00:32:46] with this all these AI stuff coming in?
[00:32:49] >> Look, first and foremost, come to terms
[00:32:50] with the fact that GRC initiatives,
[00:32:52] governance, risk, compliance are
[00:32:54] necessary, but don’t make you secure.
[00:32:56] You you have to own that. You have to be
[00:32:58] realistic with yourself. And if you’ve
[00:33:00] built an environment or inherited an
[00:33:02] environment where the security is
[00:33:04] assumed because of some compliance, then
[00:33:06] you have to break that culture. That’s
[00:33:08] on
[00:33:08] >> You’re a leader. You have to break that
[00:33:10] culture.
[00:33:10] >> Yeah.
[00:33:10] >> As a leader, you have to send the
[00:33:12] message out to the entire company that
[00:33:14] yes, compliance is important, whatever,
[00:33:16] you know, governance is important, but
[00:33:18] again, this is not this is not equate to
[00:33:20] security. We also have to pursue
[00:33:22] security or protective initiatives. And
[00:33:25] it’s a culture thing. You have to you
[00:33:26] have to modify or adjust or whatever you
[00:33:29] want to label it, but you have to create
[00:33:30] change within that organizational
[00:33:32] culture so that they understand that
[00:33:33] security has to be built into
[00:33:35] everything, not just hey, after the
[00:33:37] fact, you get this report, we’re good to
[00:33:39] go. It’s not like that. This industry
[00:33:40] made that mistake for many years. And a
[00:33:42] lot of people built tremendous careers
[00:33:44] with huge packages based on that.
[00:33:47] >> Flawed model. That’s a flawed model.
[00:33:48] >> I think it’s time to just come to terms
[00:33:50] and understand that that stuff is
[00:33:52] necessary. I’m not saying it’s not, but
[00:33:54] it’s not enough.
[00:33:55] >> Right. Sounds good. So, Andres, it’s
[00:33:57] been a pleasure having you on Tech on
[00:33:59] Inch. I really appreciate your time,
[00:34:01] your candid, the honesty behind the
[00:34:03] security loopholes, and everything. So,
[00:34:06] it was a fun conversation. I hope you
[00:34:08] enjoyed it as well.
[00:34:09] >> I did. I did. Thanks for having me on.
[00:34:13] >> [music]
