Nathanael Iversen discusses Zentera's identity-based security chambers and AI agent governance challenges
Summary
Nathanael Iversen explains how Zentera originated in Taiwan's semiconductor foundries to protect intellectual property using identity-based 'chambers' and Zero Trust Network Access (ZTNA). He highlights a case study where a client deployed 1,500 AI agents against a security estimate of 50-60, illustrating the risk of uncontrolled AI experimentation. Iversen argues that traditional sandboxing is insufficient without Layer 7 inspection and that AI agents require short-lived, dynamically provisioned credentials rather than inherited human permissions. He notes that Zentera participates in ONUG working groups to develop practical operational controls for agentic AI, distinct from high-level NIST standards. The company's R&D focuses on discovery, isolation, and connectivity, with a specific emphasis on Layer 7 inspection for AI workflows.
Episode sections
- 2:16 – 7:01
Zentera Origins and Semiconductor Security
Zentera originated in Taiwan's semiconductor foundries to protect sensitive intellectual property like chip designs and EDA files. The company uses 'chambers' to control access to these assets on an identity basis, serving major clients in the chip design space.
- 7:01 – 11:18
Zentera Technology Portfolio and OT Hardware
Zentera offers four main technology areas: secure chambers, ZTNA, OT security hardware, and AI security. For OT, they manufacture inline hardware devices that act as a security layer for embedded systems lacking native security stacks.
- 11:18 – 13:45
Overlay Architecture for Cloud and On-Prem Control
Zentera's overlay technology allows organizations to maintain security policy in a single location across cloud and on-premises environments. This simplifies management compared to scattered firewall policies and helps secure legacy systems.
- 13:45 – 16:52
ZTNA vs. VPN and Zero Trust Principles
The speaker argues that VPNs provide broad network access rather than application-specific access, which contradicts zero trust principles. Zentera's ZTNA ties access to specific 'chambers' or application bundles, ensuring users only access what they are authorized for.
- 16:52 – 18:56
AI Governance and Agent Discovery Case Study
A semiconductor client's board instructed engineering to spend unlimited AI tokens, leading to 1,500 agents running across 300 projects, far exceeding the security team's estimate of 50-60. This highlights the risk of uncontrolled AI experimentation.
- 18:56 – 24:02
Survey Findings on AI Agent Control
A survey of ~250 mid-market security leaders found that 85% believe agents can move through organizations more easily than humans. Additionally, 92% of companies with over 100 agents feel they need more control than currently available.
- 24:02 – 28:00
Layer 7 Inspection and Declared Intent Enforcement
Sandboxing agents is insufficient without Layer 7 inspection to prevent agents from spinning up off-platform processes. Zentera advocates for 'declared intent enforcement,' where agents must pass through protocol and command-level inspection within a project enclave.
- 28:00 – 36:39
Dynamic Credentials and API-Driven Workflows
Unlike humans who hold durable credentials, AI agents require short-lived, task-scoped credentials issued on demand. Zentera supports API-driven workflows to dynamically provision and remove access for agents, preventing privilege abuse.
- 36:39 – 38:49
ONUG Working Groups and Agentic AI Standards
Speaker 1 states Zentera participates in ONUG working groups focused on agentic AI control planes. He contrasts ONUG's operational controls with high-level NIST standards, noting ONUG collaborates with Fortune 500 enterprises to develop practical workflows for AI-enabled security operations.
- 38:49 – 40:43
Zentera R&D Investment Priorities
Speaker 1 outlines Zentera's R&D focus on discovering assets like AI agents, OT devices, and enterprise services. He identifies three core business pillars: discovery, chambering/isolation, and connectivity (ZTNA/chamber-to-cloud). He emphasizes Layer 7 inspection for AI workflows as a durable funding priority.
- 40:43 – 42:33
Survey Publication and Availability
Speaker 1 mentions a publication containing survey data from enterprises is available for download on the Zentera website. He describes the paper as focusing on the state of enterprise adoption rather than promoting Zentera specifically.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| AppleAAPL | equity | neutral | If you are wearing any Apple products, have one on your desktop or you know. are watching this through an Apple product, some of the designs for chips in there were done in platforms that Zentera is protecting. |
| TeslaTSLA | equity | neutral | If you drive a Tesla automobile or have ridden in an Uber recently, again, Toyota, um, you know, all of these like major consumer products, many of them have designs that were actually done um in you know chambers, secure facilities, if you will that uh Zentera is providing to the to the chip |
| NVIDIANVDA | equity | neutral | So it's it's primarily the manufacturers, the you know, Intel's, the Nvidia, AMDs of the world, uh they are the ones who are going to uh use it, right? Now |
| MicrosoftMSFT | equity | neutral | example, or you know, Microsoft gives you an agent to help you do some very particular thing inside their software stack. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
Zentera started inside Taiwan's semiconductor ecosystem to protect intellectual property like chip designs and EDA files.
Unidentified speaker (speaker_1) ·2:16
How to verify: Company website (zentera.com) or press releases detailing Zentera's founding and origins in the semiconductor industry. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera's technology is used by approximately 220-230 companies in the chip design space.
Unidentified speaker (speaker_1) ·3:26
How to verify: Zentera investor relations materials or public case studies listing client counts in the semiconductor sector. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
A semiconductor company's board instructed engineering to spend as many AI tokens as possible.
Unidentified speaker (speaker_1) ·16:52
How to verify: Zentera case study or white paper detailing the client engagement; verification via the client's public disclosures if available. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
The security team estimated 50-60 agents, but discovery revealed 1,500 agents across 300 projects.
Unidentified speaker (speaker_1) ·17:12
How to verify: Zentera case study or white paper detailing the specific discovery metrics for the unnamed semiconductor client. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera surveyed approximately 250-260 security leaders at mid-market enterprises.
Unidentified speaker (speaker_1) ·18:56
How to verify: Zentera's published survey report or white paper titled 'Agentic AI Security' or similar, available on their website. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
85% of survey respondents said an agent can move through their organization more easily than a human.
Unidentified speaker (speaker_1) ·21:19
How to verify: Zentera's published survey report or white paper; verify the specific statistic in the findings section. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
92% of respondents with over 100 agents think they need more control than they have.
Unidentified speaker (speaker_1) ·21:19
How to verify: Zentera's published survey report or white paper; verify the specific statistic in the findings section. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera makes hardware devices for OT security that sit inline between OT devices and the network.
Unidentified speaker (speaker_1) ·7:22
How to verify: Zentera product catalog or datasheets for their OT security hardware solutions. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera's technology runs on an overlay architecture to centralize policy across cloud and on-premises environments.
Unidentified speaker (speaker_1) ·11:58
How to verify: Zentera technical documentation or architecture diagrams describing their overlay network capabilities. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera advocates for 'declared intent enforcement' and Layer 7 inspection for AI agents.
Unidentified speaker (speaker_1) ·25:31
How to verify: Zentera blog posts, white papers, or product documentation discussing their approach to AI agent security and 'declared intent'. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera is part of working groups at ONUG (Open Network Users Group) focused on agentic AI control planes.
Unidentified speaker (speaker_1) ·36:39
How to verify: Check ONUG's public website or press releases for Zentera's membership or participation in working groups related to AI control planes. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
ONUG is developing operational controls and workflows for agentic AI in collaboration with Fortune 500 enterprises.
Unidentified speaker (speaker_1) ·36:59
How to verify: Review ONUG's published reports, whitepapers, or case studies listing Fortune 500 participants and specific AI control plane deliverables. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera's R&D is currently discovering 2,500 to 3,000 different AI agents.
Unidentified speaker (speaker_1) ·38:49
How to verify: Request the specific data or methodology from Zentera's published survey paper or contact their sales/engineering team for verification of this discovery count. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera's R&D covers the discovery of agents from Anthropic and other major providers.
Unidentified speaker (speaker_1) ·39:09
How to verify: Review Zentera's product documentation or technical whitepapers to confirm supported AI agent providers and discovery capabilities. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Zentera considers Layer 7 inspection for AI workflows a durable funding priority.
Unidentified speaker (speaker_1) ·39:49
How to verify: Analyze Zentera's public financial filings (if available), investor presentations, or press releases regarding R&D allocation and product roadmap priorities. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Primary sourceA survey paper containing data from organizations on the state of AI adoption is available for download on the Zentera website.
Unidentified speaker (speaker_1) ·40:43
How to verify: Visit zentera.net and search for the specific survey publication mentioned in the transcript. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
NIST has issued high-level standards around AI governance.
Unidentified speaker (speaker_1) ·36:59
How to verify: Search the NIST (National Institute of Standards and Technology) website for published AI governance frameworks or standards. For recent events, verify 2026-08-21 through 2026-11-21. Use another period only when the speaker explicitly names it.
Do your own research
- Verify the Zentera survey paper titled 'State of Enterprise AI Adoption' available on their website for primary data on the 85% and 92% statistics.
- Check ONUG (Open Networking User Group) working group publications for the specific operational controls being developed for agentic AI control planes.
- Review Zentera's technical documentation on 'declared intent enforcement' and Layer 7 inspection capabilities for AI workflows.
- Investigate the semiconductor client case study details to assess the scalability of the 1,500-agent deployment scenario.
- Look for independent security audits or third-party validations of Zentera's ZTNA and chambering technology.
- Compare Zentera's approach to NIST AI risk management frameworks to understand the gap between high-level standards and operational controls.
- Check for any conflicts of interest or sponsorships related to the podcast platform or tipsora.com AI certifications mentioned.
- Review Zentera's financial filings or press releases for R&D investment priorities and revenue breakdowns by business pillar.
Full transcriptexpand
0:00Arnand Sapatnaik: Most organizations are not behind on AI. They are behind on the thinking required to use it. I am Arnand Sapatnaik, ex-Microsoft Data and AI executive, CEO of Tipsora, and your host. This is what comes next. The podcast where we talk about building the kind of organization that actually wins in an intelligence-driven economy. Let's get into it. Our guest today is Nathaniel Iverson. Nathaniel is a chief evangelist at Zentera Systems, where he helps prospects and customers understand how segmentation, identity, and device. posture apply to securing data centers, cloud, user, and OT systems. He has over two decades of customer-facing experience with a broad background in networking, security, and virtualization. So, Nathaniel, welcome to the show.
1:07Unidentified speaker (speaker_1): Thanks so much. It's a pleasure to be with you today in the audience.
1:11Arnand Sapatnaik: Yeah. So uh you know, we talked briefly, uh, Um, you know. I I wrote a book called Feature Proof Uh Your Business, where I talk about the durable advantage cycle. Uh, it has this five steps. Find your nest, uh, diversify, build a platform, and then uh build ownership-minded team, and then invest in RD, right? So I kind of uh schedule these podcast conversations conversation around those stages. So there are going to be different sections. of the uh podcast where uh, you know, uh is is aligning to those uh stages, right? So starting with finding your news, you know, that's your introduction. Obviously like you have a you know, an educational background in electronics and you were systems engineer and now y you know, you are in Zantera and you are building uh, you know, uh security at the hardware level, right? So uh, you know, walk me through like, you know, how did it happen? How did Zantera get started? You know, where did you see the opportunity and where we are now?
2:16Unidentified speaker (speaker_1): Yeah, it's interesting. Uh Zantera's background actually comes out of the Taiwan semiconductor foundries. So our founder uh actually has a PhD in chip design. and has even taught at university in those fields. So his connections were very tightly tied to literally the largest chip manufacturers in the world. And when he started his entrepreneurial journey, one of the primary things that was a problem there was this whole area of intellectual property. You know, you can imagine if you've got the trade secrets for Apple and Samsung and everybody else sitting inside the foundry, it's very important that these companies don't get mixed up and that that data doesn't get out to anybody else. So our company actually started its journey around the idea of just access control and the idea of putting chambers around important assets. So you can think of this literally as chip designs and EDL files and some things like that that have to do with semiconductors. semiconductor manufacturing. Put the important things inside of a chamber, and then you can control on an identity basis who can get in or out of that chamber. And so you fast forward to today, and somewhere around
3:26220, 230 companies that are in the chip design space literally do that very thing every day. If you are wearing any Apple products, have one on your desktop or you know. are watching this through an Apple product, some of the designs for chips in there were done in platforms that Zentera is protecting. If you drive a Tesla automobile or have ridden in an Uber recently, again, Toyota, um, you know, all of these like major consumer products, many of them have designs that were actually done um in you know chambers, secure facilities, if you will that uh Zentera is providing to the to the chip design industry. So that's kind of where we got our start. And then as you can imagine that's now progressed into other types of business and other lines of business because it turns out that having intellectual property or things to protect is not unique to the semiconductor space.
4:20Arnand Sapatnaik: Yeah that's that's very interesting. So uh the the chamber that you're talking about, right? Is it um you know used by the chip manufacturer or is it used like after the chip is there? Uh you know, uh e you know, i e is is it put on the computer? Like where does that get used actually?
4:44Unidentified speaker (speaker_1): The chambers are being used today primarily in the design process, um, during the process of actually laying out the chips themselves and actually routing all of the billions of transistors. You can imagine this is something nobody does by hand. Uh, this is only done with super powerful computer systems and you know what's called electronic design assistance software. And so this software is so complicated Indicated that, uh, for example, say that you and I were going to make a new chip for Apple. Uh um, you know, for something that's in the iPhone, and we were going to make a new chip for it. Because that chip has never existed before, and the routing has never existed before, it's highly likely that we would actually have to collaborate with the makers of the electronic design assistance software. And so it turns out that in order to design chips today, it's a heavily interactive space. In other words, we might be in there, the maker of the software might be in there, a third party might be in there. And so you can imagine that you have incredibly sensitive data, and yet you've also got to have people get access to it. And so this this conundrum is what you know leads to the need for a very specialized and powerful solution.
5:57Arnand Sapatnaik: Yeah, yeah. That's uh interesting. So it's it's primarily the manufacturers, the you know, Intel's, the Nvidia, AMDs of the world, uh they are the ones who are going to uh use it, right? Now
6:10Unidentified speaker (UNKNOWN): Yeah, and
6:11Unidentified speaker (speaker_1): it turns out, as you would be aware, that you know that those are the big chips, right? Those are the the anchor processors, if you will. But you think about all of the things that have to do with power handling and just distribution, all of the memory controllers and all of the other kinds of chips that actually are on board besides the main um NVIDIA chips. And that's where it scales out into the hundreds of other firms, uh, some of which you would only know if you were in, you know, kind of in that industry or in that space, and yet we couldn't live our modern lives without them.
6:42Arnand Sapatnaik: Right, right. Now uh the the chamber, like that's one of the platform that Zentera provides, right? That is also U. you you have uh ZTNA and uh uh the the AI. So uh can you describe like you know uh what are the other uh platforms that are available within Zentera?
7:01Unidentified speaker (speaker_1): Yes, we we really have kind of four primary areas of technology that we serve. The first is this idea of the chamber and being able to put important things in a secure place. The second you mentioned is the ZTNA. and that's how do you get people in and out of there from as far away as a Starbucks or you know a vendor of some kind, or as close as your own employees. The interesting thing is we also have a portion of our business that is purely around OT, operational technology. For this part of our business, we actually make hardware. We actually make small inline devices that are designed to sit in between OT devices. that typically have no ability to have a security stack. They're just little embedded operating systems. And so we make a device that can basically be the security layer in front of those other devices that can't have a native security stack. And so we have that part of our business, and we have a new part of our business, which of course is AI. And I think you can imagine
8:01that in a world of agentic AI, there are still things and reasons why we need to keep some things separate from other things. And I'm sure we'll get into that. But you can tell that there's a common thread across all of this. It's you put a chamber or some kind of control around important things, and then control how those things interact with the world, interact with users and other things. So that that's the way that our business lays out.
8:24Arnand Sapatnaik: Yeah, yeah, yeah. I mean, um that there is like, you know, uh uh there is this new field uh that is evolved that is called confidential computing, right? Where where uh you know at the hardware level, you have these uh you know virtualized uh environment where uh only a dedicated uh processor and dedicated space within the processor is responsible for doing that uh computing. So that's like you know, for supersensitive uh computing, that's that's what's uh being marketed today. Um let me walk you through a day in life for me, okay? And I want to see how you know uh your platform plugs into my service. I am a fractional CTO, I specialize in cloud and AI. So uh a typical work for me is a customer uh who is in an on-prem environment and they are thinking of moving to the cloud, right? So my my first thing is okay, you need something like an express route, a dedicated VPN to connect your uh on-prem network to the cloud. Now uh once that connection is um established, then you need a harbur-spoke harbors architecture uh that connects to all the other services that you would have uh in the cloud. And then, you know, for
9:43each of the services, you might create a private endpoint. So those there is no public IP. And for each of those uh services, if there is a private endpoint, uh you create a virtual network and then you create VNet peering between those services. So that they all uh connect to each other within um that uh you know private network infrastructure. And then each of the services has its own firewall type of thing that is like network security group, right? Where you specify who can connect to it from which IP, what type of protocol and which port is available, right? On on top of that, um there might be f a firewall and gateway and all that, right? Now, let's say, like, you know. now that I have Zentera, what out of your portfolio of services that you described, what you would integrate into this plan and like you know, what type of cloud service may be replaced by what you are doing or how
10:44where it augments? Like, can you walk me through if I have Zentera Xantera and now I'm you know thinking about doing um cloud migration? wh how should I consider my uh network routing and you know uh keeping Xantera in mind?
10:58Unidentified speaker (speaker_1): Yes. Well I think that the things to think about is how many and where are the layers of complexity? Like obviously a company that's gonna go to the cloud already has whatever complexity is in their existing routed switched architecture for whatever data centers and locations. they have. And if you suddenly go to the cloud and you multiply that times all of the number of, you know, virtual pods, facilities, chambers, you know, whatever you put in the cloud, you can end up with your firewall or your fine-grained policy kind of scattered across a lot of different places. And what that the implication of that, of course, is that if you then need to change it at some time. in the future, the policy is somewhat fragile because it's in a bunch of different virtual routers, it's in a bunch of different places. And one of the advantages of our technology, because it runs on an overlay, is that you can actually have five things in a cloud and three things on the ground, put them on a chamber, and it just works. And the policy for that's in one place. And it what it leads to potentially. is a much simpler set of firewalling inside some of the cloud environments or some of the edge environments. Many of these places already have a policy that's super long, super complicated, and nobody ever wants to touch it or change it. And
12:17yet in the cloud, sometimes there's dynamic things where you do want to change the policy. You do want it to update, you know, on a faster basis. And so having one simple place that you can adjust policy and affects things in the cloud and on the ground, that can be a benefit. In some cases, organizations have access to newer tools in the cloud and they're actually very concerned about the things that are left on the ground because there's very few companies that leave their data centers completely. And they're actually looking for tighter controls on the ground. They may be happy with what they have in the cloud, but uh, they need tighter controls in the ground, in which case we'll often come in and put chambers around some of the systems that they may consider legacy. We actually have technology that goes all the way back to uh, you know, like Windows like 2007, um, you know, really old operating systems. And so sometimes there's legacy pieces that uh we just pick up and take care of for people.
13:15Arnand Sapatnaik: Yeah. So can you describe like how is ZTNA better? Like, why does VPN still exist? Is Like everybody talks about the zero trust. Whenever When we talk about cybersecurity, everybody wants to talk about zero trust. But VPN is not zero trust. So what's your solution?
13:45Unidentified speaker (speaker_1): Well, um, as most people would know, a VPN is virtual private network access, and it's the network access that gives it away. Uh, in other words, like I have my username, I log into something, and then I get access to the network. It's like I just jumped into your corporate network, plugged into a switch port, and am
14:02Unidentified speaker (UNKNOWN): now
14:02Unidentified speaker (speaker_1): sitting on some. part of your internal network. And of course, for many uses, that is a lot broader access than people actually need. In most data centers, we don't let strangers just walk up and plug into switch ports in the middle of a data center. And yet that's mostly what VPN access enables people to do. And so when we deliver ZTNA, it's very tightly bound to the idea of the chambers. chamber. In other words, if you have access to five applications or 10 applications or 15 or whatever the case might be, we're going to go ahead and associate your username with those particular chambers, and you will be able to get into those chambers only, and you will not be able to browse the network. You will not be able to do all of the other things if you're headed into the cloud. You're going gonna to go into the cloud and have access to the three or five or seven things. that you're supposed to have access to. And so in this way, the access is always tied to a username and that username is always tied to a discrete set of kind of services or application bundles instead of a very broad just network context where you often lack And granularity.
15:10Arnand Sapatnaik: Yeah, yeah. That's great. So uh you know the last piece of the puzzle that there is AI. Like people talk about AI and AI safety, but the end of the day, AI is an application, right? So you already have your network security and firewall, your data governance if you have implemented correctly. So that is data security, and that is you know role-based access control uh for everything. So
15:36Unidentified speaker (UNKNOWN): it
15:37Arnand Sapatnaik: that that should essentially take care. of AI. You know, um, you can have your safety system prompts on top of it. Is there anything else that you recommend, you know, your your NSAID AI? Like what additional thing does that do?
15:53Unidentified speaker (speaker_1): Well, I think it's interesting. You mentioned something important, which is that people have governance already. And the governance controls things like data and users and access and all of the these kind of things. And it would be interesting if most people actually took all of those beliefs and policies and applied them to AI. And maybe it sounds silly, but that's actually a point that a lot of organizations are running past their own controls. They're running past their own governance. In other words, you know, there's people experimenting with AI out. outside of those governance frameworks. And I think that that's probably one of the biggest risks that many organizations face, they're not treating it just like another software package. They're not treating it like something that we have to just proof of concept, test in a limited way, roll out in a very defined way. Um, there are many organizations that are just turning developers loose with huge. amounts of token buckets and saying go make things. So they are. And so it's in that world that we actually a year ago had a pretty big wake up call. One of our big semiconductor companies in Taiwan actually came to us and said, our board of directors told engineering, spend as many tokens as you can, we've got the money. And so you can
17:12imagine what an engineering group got told to do What did when they got told go play with technology, right? They went and played with technology. And the security team came to us and said, Well, you know, we think we've got about 50 or 60 agents running inside engineering. We'd like you to, you know, run some discovery and then help us, you know, kind of get some control around that. Well, we did the discovery. They had 1,500 agents running across 300 projects. And so I think that this kind of thing is a thing that happens very often. And that led to us then starting to ask a whole host of questions around how do you discover the all the agents you have running? How do you begin to put sandboxes or some control around the agent? What are the extra controls that you need that aren't in a firewall, that aren't in just a typical network segmentation policy? And it led us down this journey that we've been on for the last year to try to help organizations both to understand and control agentic AI. So I've left kind of a number of threads open there, but I think it can open into an interesting conversation if you'd like it to.
18:21Arnand Sapatnaik: Yeah, yeah, definitely. I mean it's it's an ever evolving uh area. So that there is that there is already like, you know, agent control panel that we people are building um, you know, to keep tabs on the agents, uh right. So yeah, I mean uh that that's an interesting area that that that continues to evolve. Um you recently published a uh paper on uh a survey that you did with uh you know cybersecurity uh leaders. So can you share like what that study was and what was your finding from that?
18:56Unidentified speaker (speaker_1): Yes. Uh that paper is fairly interesting. It's went out and surveyed about 250, 260. Some people didn't respond to all the questions. So you know you get these little variable this little variability in the data. But these were primarily uh directors through C-level executives at what I'll call mid-level companies. Now, a mid-level company being somebody that's multi-billions of dollars and uh, you know, kind of just outside the Fortune 500. In other words, there weren't a lot of Fortune 500 respondents, but you know, maybe down through the Russell 2000, uh, that level of company. So in here were manufacturing companies, super large healthcare consortiums, uh, you know, various technology, even banking and you know, retail. and finance kinds of companies. So it was fairly diverse. And when we interviewed these people, we wanted to get a state of the union for kind of how they were thinking about AI, AI governance, the big problems they were facing, and things like that. And we found some very interesting uh data points. Number one is that
19:56most people are you know experimenting with AI, not a surprise. There's a lot of people that are almost everybody everybody's doing some kind of co-pilot or like user assist kind of AI. Um, the agentic use uh tends to be it tends to start with vendor AI. So Salesforce gives you an agent, for example, or you know, Microsoft gives you an agent to help you do some very particular thing inside their software stack. And so those don't feel super agentic because they're not really taking out. action outside of a defined workflow. Like they have very strict bounds from the manufacturer. But obviously as you get into tech, you're getting a lot of people that are using AI for various coding things. You get into manufacture, there's a ton of data that they want agentic AI to try to look at for optimizing lines, production schedules, seasonality and retail you know, fraud and banking, you know, all of these kind of things. Some of the interesting findings that we had,
20:59we learned that if somebody has more than 50 agents, they uh about 80% of the companies think that they need more control than they have. If someone has over 100 a hundred agents, like 92 of the respondents. think they need more control than they have. So it's kind of interesting because what it says is that the more agents you have, the more you realize that you need an entirely different level of control than kind of the router switches, firewalls that you've had for a zillion years. And so that was a very interesting finding. Another was that 85 eighty-five of respondents said that an 80%. agent could move through projects and move through their organization easier than a human. So that's kind of a very powerful statement. You fire off an agent to do one thing, and it's able to move around your environment easier than a human could. And so the implications of that, of course, I think we've seen in some of these early, you know, AI-based AI based attacks where AIs are able to come in and agents are able to execute thousands of commands in 30 minutes or less in a way that even a human hacker would be completely unable to do with the best automation. So it's interesting some of the things that kind of came out of it that point
22:18towards usage, lack of control, you know, a very different operating uh modality, like maybe. we would observe that the agents aren't reasoning like a human, but they do reason. And so that means that their behavior patterns are perhaps different than a human. And so some of that data came out strongly in the survey.
22:39Arnand Sapatnaik: Yeah, yeah. people hype it up that this AI went rogue and did this. So this is becoming a very visible threat to the point that now no like you know it it's becoming top priority for the government. So uh all the tech giants have to get together and find a solution to this. Um so l let's uh move to the uh next aspect of it, right? Now your company started out uh you know as as a uh you know hardware chip manufacturing, right? You know, I would say like similar to Nvidia, like Nvidia started as with the GPU, but then they became this AI platform or software. So what are some other areas like that you are thinking about? Like you you do have like you know, you have your chip, uh you have your uh hardware level security, you are also doing uh networking, now you are doing a AI, you are doing application security. So you are kind of you know diversifying within the cybersecurity, but what are
23:45some other areas that you are exploring uh to kind of diversify and you know and and kind of uh you know go into more uh emerging areas?
23:55Unidentified speaker (speaker_1): Well, one of the things that we learned uh in working with our our customers and in kind of the early agentic days, you
24:02Unidentified speaker (UNKNOWN): know,
24:02Unidentified speaker (speaker_1): maybe nine, 10 months ago, is we learned that it was one thing to put a sandbox around an agent. In other words, like if you can make the agent run inside of a combined space, then theoretically you can control exactly what the agent is doing and where it's going. And that's certainly true from a network perspective. Once you get the right shell around it, you can have some control about where it goes. But what we found is that that was not enough. If you just restricted an agent to only go certain places, it could probably figure out how to use those destinations or those LLMs to spin up other processes and to do other things, you know, kind of off box or off-platform. It didn't really feel very restrictive for it to just be able to get to, you know, some big LLMs. And so what we learned. is that it was very important to have a layer seven control, if you will, or a gateway or a proxy. Like these are all words that get at the idea of you need to have your sandbox actually connected to something that can actually inspect the contents of what that agent's trying to do. And if you make it so that the agent has no
25:11way to go except to go through that layer seven content. control or that gateway, uh then you have the ability to actually really lock down what the the AI is capable of doing. And I think there's a new kind of set of language starting to emerge in the space around like declared intent enforcement. This idea that you know an agent shouldn't be allowed to just invent everything it wants to do. We don't allow humans to do this. in enterprises. Like if you're in marketing, you don't just get to check code into the code repository, right? And so in a same way, you know, we have a belief and our customers have a belief that agents need to be put into some kind of a project boundary. We call this an enclave here at Zentera, but this idea that you're gonna going to need some agents, you're gonna going to need some LLMs, some MCP servers, probably some back back-end end data sources, maybe some. industrial automation, whatever the case might be for somebody's business. But these things exist inside of a project context. And so how do you establish that project context and then get the right controls so that the agents kind of have to pass through protocol-level inspection, command-level Command level inspection, you know, file system access level inspection before that? they actually take access inside that broader project. So a lot
26:31of our area and focus is in that right now, trying to provide the controls that organizations need to keep increasingly capable agents confined to the projects that they belong in.
26:42Unidentified speaker (UNKNOWN): Yeah,
26:42Arnand Sapatnaik: yeah. I think, you know, like what I see is that you're kind of evolving into being that, you know, inspect every everything, you know, uh put a cell around everything, uh, you know, have a centralized uh policy. Uh you know, so that's kind of your thing, and you are able to apply that starting from the chip to networking to now AI, you know, and and of of course applications, right? So um I I I think that that what I would consider a good platform because when I talk about uh you know uh platform is that that's your core competency. And then you have the different branches where you are able to apply. So I think, you know, uh uh ha having that centralized governance and creating cell, I think that's where your core expertise and you are able to apply that to hardware, networking, application and now AI. Right? So so uh that leads me to the uh next thing, which is you know becoming the platform, right? So if
27:43Unidentified speaker (UNKNOWN): if
27:43Arnand Sapatnaik: if that's your platform, right? So the platform means that others should be able to build on it, right? So what are some ways that other software companies or you know uh other vendors can build on top of like what you you're are building?
28:00Unidentified speaker (speaker_1): Well, as you would imagine, this kind of capability can be fully automated via API. And that becomes particularly important, we believe, especially as we're getting into this agentic world. You know, we have a habit of issuing fairly durable credentials to humans. Like I, you know, am a new employee at your company, you're going to issue me a laptop and some credentials and get give some user logins. to applications and all the things that your onboarding process does, right? So as a human, I get that. And it really doesn't get taken away until I leave, or maybe I change roles. And so some of my role-based access changes. Um, these changes are fairly slow. But what we're seeing is that, you know, you don't really want to give agents that kind of, you know, control. You don't really want an agent running with my user permission. Think about it. Like this agent may have access to go do things, but if it runs with my user permissions, then that agent actually has access to my payroll data. It has access to my email to send nasty grams to the CEL, like it could do whatever it wants. And so this kind of situation starts to imply that identity is very important for agents. I think you mentioned this even early in the talk, but it turns out that identity needs to have some different characteristics. It probably needs to be granted on
29:19demand. It probably shouldn't exist all the time. It should probably exist for very narrow tasks. In other words, somebody gets, you know, five minutes of access to do a thing, or they get 15 minutes of access to do a thing. And then that credential goes away and it has to be asked for again. And so we're finding that the credentials need to be held. off box, they need to be time limited, they need to be, you know, very actively scoped. And when you start doing these kinds of things, then you start talking about API-driven workflows, right? Where if you're going to have layers of automation, that automation may need to be able to interact in order to get the right tokens or the right access credentials, you know, which give you access to a thing which is dynamically provisioned. and then dynamically removed. And so we see ourselves as, you know, ultimately operating inside of that larger context of a modern automated enterprise.
30:14Arnand Sapatnaik: Yeah, yeah. That's great. Like I I I I would compare this to like, you know, what what Stipe and others did uh i i i in the financial world, like, you know, you don't have to kind of uh uh there are these APIs but for open banking. So, you know, if you are building a financial application and you use that API, uh then like you know, you don't need to reinvent the wheel. So similarly, like what you are doing, if I build an A AI application or I build an API that I want to sell to customer and I'm integrating, you know, with with your API, then then you know I don't need to reinvent the security aspect of it. I can say that, you know. these guys have done it. I'm integrating that. So it's it it's it's as secure as it can be, right? So then I'm able to kind of, you know, add um and simply by using your API, I'm taking care of the security part. And I am able to um add value in terms of functionality in my API, right? So so that's how it like you can could have an ecosystem um of people who are
31:14like, you know, okay, they will take care of security, we'll take care of functionality.
31:19Unidentified speaker (speaker_1): Yes. You know, we've also had um some organizations that manufacture OT hardware. So they actually make PLCs and industrial controls and temperature sensors and feed rate controllers and you know all of these kind of things that are part of modern industrial automation. Uh we've had some people ask us, you know, can can we OEM or bundle your uh chambers for OT? Uh, because you know, we don't want to be part of one of these blowback things where people hack our controllers and then you know they're in the news, uh, you know, even though they're just the manufacturer of the OT. And this gets to your idea of like it's a platform, it's a play that you know, somebody who needs to be able to isolate OT or get just their technicians access to the OT and no one else. Like these things aren't exposed. The wastewater treatment plant controllers are not just exposed on the internet. The refrigeration controllers are not just exposed on the internet. Um, you know, these kind of things we're seeing it also play out even in the physical world. So, you know, the world of APIs is one thing, but then it's also like this additional layer that can be added around some of the OT devices that traditionally have not had these kind of you know protections.
32:33Arnand Sapatnaik: Yeah, yeah, that's great. Like O OEM is already an established. So you know if if you can kind of sell your uh you know hardware uh um infrastructure as as an OEM, then that others can build their devices and their software using. yours so that that uh uh uh opens up another ecosystem. Um now l let let's get into the the people management aspect of it. Like uh you know like y you have people in Taiwan here so uh globally like you know th there are different roles. So how how are the people incentives uh you know created within the company to make sure that you know people are aligned to the company bottom line.
33:18Unidentified speaker (speaker_1): Well, um I think that like most uh you know high technology companies, there's a pretty tight binding between our customers, our engineering department, and then what ultimately happens inside of our sales and marketing organizations. The primary flow of information you can imagine that. we've our history has had us had very close interactions with some of the EDA vendors and various design shops and other things. So our history as a company is one of being very embedded with our customers, not just in the sense that our technology is deployed there, but in the sense that we're really working to understand the deepest parts of their application flow and make sure that our solutions. are customized to that environment. So our whole product organization and engineering is very focused towards that practicality. And what happens out of that, of course, is then we get these strong use cases. Um, you know, it's one thing to say, OT security, AI security, um, you know, and these are like big giant buzzwords that mean almost nothing by themselves. But you get into the very specific. thing of you know hosting a chamber for a design session, and all of a sudden then you have a very powerful case study that our marketing or our sales department
34:37can use to say this is what it means. This is how it actually comes out into the marketplace. And so I think, you know, for those on our executive staff, you know, we're trying to constantly make sure. that we're operating as one and you know, bringing a unified force both to the market and to our customers.
34:58Arnand Sapatnaik: Yeah. And and like, you know, how much freedom do people have? Like if somebody has an idea of building a new product or uh you know enhancing existing process or product, uh what what freedom do they have to get kind of you know create a new project initiative?
35:14Unidentified speaker (speaker_1): Well, we're at a phase in our company where good new ideas are always welcome. We do not have a committee of no. We're in a fairly flat environment. Uh it is an environment in which anyone with an idea can bring it forward and people do. And you know, we're just trying to be the company where the best ideas win. Uh that is something that is easy to say and is much harder to do. Uh but we're still small enough that we don't have enough layers of politics uh damping down all the good ideas, ensuring that only the mediocre rise to the top. We're still at a place where good ideas can win. So I'm personally happy about that.
35:51Arnand Sapatnaik: Yeah. So you know I another thing about uh you know um uh sustainability is that you know besides taking care of your own people, your own product, you have to also uh manage the external environment, right? Which is like you know, the policymakers, the the the market, you know, the re regional market, right? So are you part of any governing uh body or any uh of these you know um uh committees or organizations that set the standards around uh security or hardware manufacturing where uh you know you are able to kind of you know uh make a case for you know how the those policies should be saved and how those standards should be created.
36:39Unidentified speaker (speaker_1): Yes, we're actually a part of a number of working groups at ONUG, the open network users group. Uh I believe that they're doing the strongest work on agentic AI right now. from an overall control plane perspective. Clearly, there's places like NIST and others that have started to put some standards out around AI governance and they're fairly high level. What we love about what ONUG is doing is that they're actually taking those kind of standards and then developing real operational controls and workflows, working with a lot of the largest enterprises, you know. the Fortune 500 specifically. And so we're involved in several of the working groups, setting policy around agentic AI control planes, how to build uh, you know, AI-enabled SOC and NOx and you know, things like that. And so uh that's a place that both we're contributing from a vendor perspective, and it's also a place that we're learning uh from the enterprises. You know, this is a time in this space.
37:39where it is actually highly collaborative at the top levels of the game. And everybody is trying to figure this out together to get the answers, you know, on the customer side to run their organizations and on the vendor side for us to know exactly what to build that's useful. Like not just something that a bunch of PhDs thought about, but That actually code that matters and that can run in high-stakes environments to protect. critical assets.
38:05Arnand Sapatnaik: Yeah, yeah. Now, um, you know, moving to the next stage, uh, you know, you you uh allow people to create, uh have the freedom to create something new. Um, you know, this is an ever-evolving feed, and we are at a stage where you know uh new invention, like new models and new capabilities are coming faster than you know oh we we can learn about it. You know, what type of research and development, you know, uh investments are happening or I would say like you know if you had unlimited funding, w where would you want to invest in research and development to create that next best thing?
38:49Unidentified speaker (speaker_1): Well for us here at Zentera, our eyes are always on a couple of things. One of them is on the discovery of assets that needs to be protected. So you can imagine that this covers everything from AI agents. We have part of our RD that I think we're currently discovering, you know, 2,500, 3,000 different agents that are most of them, of course, well outside the big ones, you know, from Enthropic. Anthropic and, you know, other places. But, you know, we have a constantly ongoing thing there. We have a constantly ongoing thing about recognizing OT devices and different types of enterprise services. So there's a part of our business that will always have a discovery component to it. There is a part of our business that's always around chambering and isolating. That is, of you know, kind of trying to keep important things separate from other things. And there's a portion of our business. that's then about connectivity. And that's whether that's ZTNA or it's chamber to chamber, or it's chamber to cloud, chamber to you know to AI. Uh, I know that we would also uh at this point moving forward,
39:49we're probably always going to have something invested in layer seven as it relates to AI and agentic workflows. So Uh whether you call that a gateway, a proxy, uh, you know, a session controller, people use different words. Um but that idea is one that I think is a durable funding priority inside our company.
40:08Arnand Sapatnaik: Yeah, yeah, definitely. Um especially like in the AI world, the session management and and making sure the context is limited to that conversation, not shared with other user, right? Uh that that's that's a hot topic. This cannot emphasize enough on that. Um are there anything, you know, like you had your uh you know uh publication around uh that survey that you did. Um are there any new products coming? Are there any new announcements uh that you want to make uh on the show?
40:43Unidentified speaker (speaker_1): Probably not on the show, but as you can imagine, uh you know, Santa's elves are always busy in the workshop. So um you know we're we're constantly. building new things. With regards to that paper, what I will say is that if people want to get that paper, it is available uh with all of the data from these organizations, it's quite well worth reading, I think. Uh, and so that's available for download. Uh you can certainly get it off of the zentera.net website, or maybe it'll be linked in the show um, you know, information. But that's something that's available to all of the the viewers of the podcast. And uh that's not about Zentera, it's it's really actually more about enterprises and the state of adoption. So if that's interesting to you, definitely pick it
41:25Unidentified speaker (UNKNOWN): up.
41:26Arnand Sapatnaik: Yeah, yeah, definitely. Okay, so that uh takes us to the end of our show. If this episode shifted how you think about AI in your organization, send it to one person who needs to hear it. And if you're you are ready to add, act, visit tipsora.com to get certified or reach me directly at arvancipattack.com. The intelligence economy is already here. The question is whether you are building for it. That's a wrap on today's episode of What Comes Next. If this conversation gave you a new way to think about AI strategy, share it with someone who needs it. You can find everything I'm building at tipsora.com tipsura.com, in including AI certifications for you and your team. Connect with me on LinkedIn by looking up my name, Arnan Chapatnaik. You can find Nathaniel on LinkedIn or visit his company site, zentera.com. Until next time, build the architecture. The advantage follows. So, Nathaniel, thanks again for being on the show. It was great talking to you. Hope to keep in touch.
42:25Unidentified speaker (UNKNOWN): Great.
42:26Unidentified speaker (speaker_1): Thanks so much.