DataTribe Challenge Finalist: Tensor Machines

Welcome to the Cybersecurity Go to Market podcast for our special showcase episode where we're talking to leaders of the company selected for the 2025 Data Tribe Challenge, where Data Tribe, the Cybersecurity Foundry, have identified five startups as being the most innovative in cybersecurity today. I am your host, Andrew Monahan, and today we have Muneib Rasul, co-founder at Tensor Machines. Muneeb, welcome to the podcast.

Thank you very much, Andrew.

Looking forward to this episode. It's always cool talking to people doing really interesting things. So uh with a with a lens, of course, from the go-to-market or the business side, we're not the audience is not usually super technical, Muneb. So keep that in mind as we're going through this. But let's start at the beginning. What's the story behind you yourself first getting into cybersecurity?

Yeah, talking about non-technical audience. I didn't start in cybersecurity or in cyber for that matter. I was actually an econ major during my undergrad, but I built a small social networking app that used game theoretic principles and match students based on their interests. So it kind of took off, and that experience pulled me full further into the world of uh product development. And my curiosity evolved into a deeper fascination with how technology impacts trust in society. So as I grew my capacity to build products and work on more complex problems, my focus moved closer to the intersection of hardware and software. I was a product lead at a startup in connected mobility. We were building these vehicle telemetry systems that collected driving data through OBD sensors to personalize insurance. It was one of the most technical, technically elegant products I'd worked on. And it really did reduce the insurance price and cost for drivers. And yet, whenever I tried to sell that product to one of my more cost-conscious friends, the focus of the conversation turned into trust versus cheap insurance. They were uncomfortable with a system that could see everything about their movement. And it was my first realization that complexity, no matter how brilliantly it encases the technology, it collapses if people cannot trust what it touches. So from there on, I moved into hardware at Meta. I worked on mixed reality devices, which are amongst the most sensor-dense consumer products ever built. We keep we kept adding with every device iteration more gyroscopes, cameras, other sensor modalities. And as the system got smarter, because it learned more about the world through all these sensors, it also became more fragile. Every new sensor adds another interface that, though, makes the experience more immersive, but it can also be attacked. And that was around the same time when the generative AI and its ability to generate information, I saw firsthand at Meta that how easily this kind of misinformation or fake information spread across social media could destabilize whole systems of belief. And that was kind of a triggering point of my segue or my step into the cybersecurity and start thinking about things from that perspective. Because if fake information can corrode social contracts, what happens when fake sensor data start staring machines? So the question stayed with me, and I was able to connect that dot back to my experience as a cargo specialist when I was in the US Army, moving equipment where precision and timing were everything. And if autonomy is going to be the centerpiece of efficiency in our mission critical world, then a broken link will deeply jeopardize that mission. As the UAVs and autonomous robots take over the world, my deep concern is about the operational integrity of autonomy. So I got into cybersecurity because I'm concerned as embedded systems are scaling and interacting more deeply with the real world, their interface density is exploding and their trustworthiness is declining.

Well, I think there's interesting points in there for me personally. The first example you give about the insurance side, I've got two teenage daughters, both driving, and you get offered that, right? And I've got to admit, I said no. And partly because of what would happen to the data, but partly because I honestly feel like the insurance industry has got a bad track record of using using things against you to talk about claims or not claims or accidents or not accidents. You know, I think it seems like everyone everyone in my circle has a story somewhere about you know asking a question about an incident and then getting hit for having an accident, things like that. So I think I guess in many ways it's the use of that telemetry, the use of that data is what can erode trust as well, right?

Yes, yes, yes. So when we think about cybersecurity, traditionally that meant focusing on the software and focusing on the software data. But the physical world is increasingly being run by embedded systems. So now the world needs a security model that understands hardware and physical systems and their physics. So that also kind of, as you mentioned, like what's the source of the data and how do we protect that new sources that we are creating became the seed of me building not only for security, but also the resilience of these uh cyber physical systems.

So let's go back in time a little bit to the actual genesis of Tensor Machines. And well, yeah, you're a co-founder. So where were you in in time and place? And where who were you with where you could have sat there or stood there or walked there and said, you know what? I think we should start a company doing this. Where was that moment in time?

Yeah, yeah. The founding moment came while I was still at Meta. I was working on AI-driven devices, and uh, Meta's research culture is extraordinary. Its large language models and perception AI are setting the global benchmarks, but I am wired to think in counterweights. I ask myself that as this intelligence layer right in front of my eyes getting stronger, what's the balancing layer that is gonna keep it honest? So every market needs a stabilizing metric, a countermeasure that keeps growth aligned with the integrity. So around that time, with that concern in my mind, I reconnected with my friend who is my co-founder now, Professor Hafiz Mullik, who teaches computer and electrical engineering at the University of Michigan. He has worked with uh DOD and many of the cyber physical companies. So I wanted to learn more about his research on signal authentication because he has been featured actually for detecting deepfakes through his through his more electrical engineering approach during the last elections. So, and he has also done a lot of work on hardware forensics, sensor data integrity. It is heavily cited. So I remember him visiting me in Menlo Park, and we spent the whole night talking ideas together, especially being concerned about the trust's cascading impact in an increasingly autonomous world. He told me that if falsified media erodes social trust, then what will those falsified sensor streams uh will do? They will erode operational trust. And I kind of agreed with him because I told him that when sensors are manipulated, the consequence is physical. Because when the media is manipulated, it's political consequence. And then the attack vector moves from persuasion, as we have seen, that foreign actors have done on our social media to actual control. So we then started working on this idea together. I'm, of course, coming from the product and I have done a lot of like uh sensor-based consumer hardware, and he has his deep background in signal forensics. So there were no per se any napkins involved, but we stared at a lot of oscilloscopes, we ran waveforms through classifiers, and we drank far too much coffee. And at the heart of it all, we asked, like, can we make a machine prove its own identity while it's running? And that's how kind of Tensor was born.

Well, that's that's cool. You've given us the the kind of background of why you got to this point, some of the relevance. Let's try and make it a little bit more real, though. Who is the person, who is the type of company that might have this problem? And what exact problem do they have? They're sitting there right now going, I can't do this or I can't see that. And I wish there was something that married hardware and software and gave me the telemetry around that.

Yeah. So for the question of who we have defense and aerospace integrators, they are running UAS, UAS or UXV type of systems and sensor-dependent payloads. We have automotive and industrial robotics, OEMs, tier ones, they are deploying L3, L4 level vehicles and robotic cells. We have computer infrastructure operators like GPU clusters, where board health and provenance directly affect the throughput. But the problem is that autonomy is scaling faster than its ability to stay reliable. Safety critical devices from sensors to ECUs, robots, drones, can be spoofed, swapped, degraded without any continuous verification. And right now we mostly have software-only controls that act after the bad data has entered the control loop. So, for an example, in defense, GPS spoofing has crashed fleets of drones in Ukraine and Black Sea. Counterfeit chips slipped into our DOD aircraft supply chains and have forced multimillion dollar recalls. So these incidents kind of are revelation of how hardware, because it's not the first-class citizens in the cybersecurity, can cascade into systemic failures. And the commercial world is kind of next from defense to commercial. Billions of new sensors are being deployed across vehicles, factories, and data centers. Each one is an unverified interface and an opening where bad faith actors can weaponize this modern infrastructure.

Is it the fact they just have no way to do it right now, or there's no way to do it cost effectively, or there's no way to do it in real time? What's the real problem here?

It's it's kind of all of the above. There is the first of all, we are just starting to come up with these hardware-first approaches, and it's very hard to do it runtime. And there's a very high risk of an attack on the modern hardware because the hardware and its internal networks were not designed for connectivity and perception. Hardware was assumed as a self-contained isolated unit when networks, when the network standards like a can were set. So all the various attack vectors on the sensor and connected hardware, they remain highly uh feasible, especially at the runtime. And then they remain undetected because we just have not been thinking from the physics perspective and building the right telemetry because of our software first detection approach until the system surface. So with Tensor, we do two things. One, we make it harder to attack an autonomous system. And then we raise the operator's ability to control and detect each attack from hardware first perspective. So this operational risk of deploying such a system decreases significantly.

Now, is it the general approach that you have, which is innovative, or have you got some secret sauce underneath the covers that you've developed that makes all this possible?

We have about five patents on how we do it. First, we are providing a runtime physics-based verification for autonomous systems to deliver operational integrity continuously. So, again, for the non-technical audience, think of it as a heartbeat monitor. But for machines, so if the rhythm changes, you know the system's been tampered with. And we do this at two levels. At the hardware-based level, we continuously scan analog side channels like electromagnetic interference, current transients, thermal map, uh, thermal heat map of the hardware. And then we use a proprietary built from scratch physics-grounding machine, uh physics-grounded machine learning model that analyzes and learns the behavior of these analog signals. And then it generates a unique baseline for each device that it is monitoring to continuously check for security, identity, and also for the resilience and the health of the device. And then at the second level, at the perception level, we protect the data path. We deploy specially trained agents at firmware level, which monitor every data packet that the sensor is acquiring from the real world. So any manipulation is isolated before it hits control. And these agents are not encrypting the data. And instead, we use special temporal and spatial methods to detect changes. And this makes the whole thing extremely compute efficient.

So going back in my experiences over the years, you use some terminology there, which I sort of kind of recognize packet inspection is one, right? I mean, that network IPS IDS has been around for a while. I heard kind of behavioral analysis stuff, like you could take a baseline, and then if anything moves off the baseline, then you can say that something's not going quite right. So you're you're taking some of the sounds like the concepts that have been around for a while, but figuring out how to do it in a very different environment that is critical for one, real time for another, right? Like you can't be storing and forwarding these things, and doing it in a manner that sounds like both the integrity but also the securities at play here. Am I on the wrong along the right lines there?

Yeah, absolutely. So the concepts are there, but then the scenario, the sampling rates, and how do we acquire these signals from machine to machine, interface to interface? And then our machine learning models are the ones that are making the assessments. So it is scalable and it recognizes uh the continuous evolving landscape of various attack vectors is is our secret sauce.

Yeah. Sounds incredible. Let me switch gears on you a little bit. You're very early stage, right? You're still in stealth mode. You're building up the technology side of what you're doing right now. And imagine you're you've got one eye on the future. Let me ask you a kind of existential question that seems to uh be coming around for a while now. What's more important, what's most important to a startup and maybe in cyber in general? Is it product or is it distribution?

That is a very good question. So I think for us, the product is the moat. All the patents that we have acquired, all the research and development we have done from NSF has given us many grants to build this. But the distribution is really the multiplier. Because if we don't have a good distribution, then all the the technology will remain isolated. So I'm a huge fan of distribution. So that's why our pilots run through the partners that are already embedded in defense and industrial supply chains. We have Digi Technologies, LG LHP Engineering, who are into embedded system diagnostics. We also validated our sensor protection technology by completing a pilot with Ford on sensor authentication. And we are currently doing a POC with Carnegie Mellon Robotics on drone payload verification. We designed our product for retrofit, and that makes distribution inherently easier. So adoption kind of can scale naturally through integrators that are already serving the market. And our product can be deployed without redesigning any hardware.

And it feels like going forward, your ideal go-to-market is going to be through OEMs, right? It's going to be through the either the people that use the embedded systems or the embedded systems providers themselves. Is that how you're thinking about it?

That's right. Yes. Especially in the robotics and autonomous world, we are looking at integrators who implement these systems.

And as you look forward to 2026, Munip, what's in store for Tensor Machines?

So our mission and the vision is to make operational integrity a measurable standard across autonomy. So first is that we are going to deliver a productized version of our technology. We are calling it TensorStack version one. This is the first full stacked runtime protection for autonomous hardware. We are going to expand the POCs. Right now we are looking at drones and vehicles, but I think the need is all the same with GPU grids. They are extremely vulnerable. So to data centers and also to industrial robotics. Right now we are part of plug and plays defense and aerospace accelerator. And through that, we have the opportunity to run real-world large-scale pilots through Texas AM's Bush Combat Development Complex. And on the technology side, we are adding new physical modalities like optics and deeper thermal imaging to our core machine learning model that learns the physics of the hardware. So the really the goal in 2026 is to establish an understanding in the market that just as uptime defines the economics of cloud computing, operational integrity will define the economics of autonomy, and Tensor is building that standard.

Well, that's a great way to wrap this up, and thanks so much for joining us. And uh good luck in the Data Tribe Challenge and also for 2026.

Of course. Thank you for having me.