Israeli Cybersecurity, Who Can You Trust Now? : My takeaways from I3’s Webinar
Find the full recording of the webinar here: https://i3-capital.com/webinars
Reading back through our recent cybersecurity panel, I kept coming back to one question: what, exactly, are we asking people to trust? Cyabra, HEQA and Lumin AI Security approached that question from different directions. One focused on the people and narratives we encounter online, another on the links carrying sensitive data and the third on the AI models inside an organization.
Three companies. Three different problems. But the discussion made a useful distinction: protecting access to a system does not, by itself, establish that we can trust the information, infrastructure or intelligence inside it.
I think of this as the trust stack. It gives us a way to connect disinformation, quantum security and AI model protection without pretending that they all belong in the same product category. They do not.
The trust stack
Let’s set the table. Cyabra asks whether an online account or narrative is authentic activity. HEQA asks how to protect the links between critical systems. Lumin AI Security asks what happens when the model an organization brings into its own infrastructure becomes a source of risk.
These are separate buying decisions, not three features of a universal cyber platform. That matters. A compelling technical story still needs a customer who owns the problem, has a budget and can deploy the product.
For Israeli founders, that is the commercial question I would put ahead of the usual debate about how large the cybersecurity market might become. Where does your product sit in the trust stack? And who needs it badly enough to pay?
Cyabra: Who is behind the noise?
Cyabra starts with the public conversation. In the panel, the company described how it connects accounts, behavior and amplified content to help governments and brands distinguish authentic activity from coordinated manipulation. The question goes beyond whether a particular post contains false information. Who is pushing it, and do those accounts act together?
Its work around Elon Musk’s Twitter acquisition offers a concrete example: Cyabra analyzed accounts for Musk’s team during the dispute over bots on the platform. That engagement put the authenticity question inside a consequential business decision.
The commercial model sounds familiar. Cyabra described recurring subscriptions, usage based packages and opportunities to expand customer accounts. But the underlying purchase is less familiar than an endpoint security license: customers need to recognize manipulation early enough to do something about it.
My takeaway is that detection only earns its keep when it changes a decision. Which narrative deserves attention? Which accounts warrant investigation? When should a brand respond, and when would responding simply give the campaign more oxygen? That is where I would look for evidence of customer value.
HEQA: Trust the link, not just the endpoint
HEQA moved the discussion below the application layer. Its panel presentation focused on the connections between data centers, clouds and other critical systems, arguing that organizations need visibility into the physical link as well as protection for the data traveling across it. An identity check cannot answer every question about that path.
HEQA describes its product portfolio as combining quantum key distribution, integrated key management and a post quantum cryptography overlay. The distinction matters: the company presents more than a software migration to new cryptographic algorithms.
During the panel, HEQA also described capabilities for detecting interception on a link, stopping key delivery before exposure and diverting traffic. Those are company claims. As a potential investor, I would want to understand the deployment conditions, the operational tradeoffs and how the customer tests that protection.
The company presented its technology as operating in production rather than remaining a lab exercise. Good news to hear. In my opinion, for a carrier or an infrastructure operator, the test is not whether quantum security sounds urgent; it is whether the system can protect a critical connection without creating a different operational problem.
Our view: do not let a prediction about the arrival of a powerful quantum computer substitute for a deployment case. Ask which links need protection, how long the data must remain confidential and what the customer can implement now. The calendar matters, but so does the network.
Lumin AI Security: Owning the model means owning the risk
Lumin AI Security brought the discussion inside the model. The company focused on organizations that deploy open weight models within their own infrastructure rather than relying exclusively on a commercial AI service. It pointed to cost, privacy, sovereignty and the ability to tailor performance as reasons for that choice. Gilad easily explained the difference between traditional infrastructure and his solution in plain language that resonates with everyday investors.
He explained how although more control sounds attractive, it also creates work. In the panel, Lumin described a security process that starts with examining the model before deployment or fine tuning, continues through runtime protection and extends into governance.
The company calls its approach mechanistic interpretability: examining the model’s internal workings rather than relying only on controls around its inputs and outputs. Lumin said it uses that visibility to identify manipulation, safety concerns and vulnerabilities. That is the part of the presentation I would want to examine most closely.
Lumin also described a lightweight protection layer suitable for deployment within customer infrastructure and on edge devices. I would treat that as a proposition to test, not a reason to skip testing. Which model, which attack and what performance cost? Comparisons need a benchmark before they mean much.
My takeaway: bringing the model in house does not finish the sovereignty discussion. It starts the responsibility discussion. If you control the intelligence, you need to understand how you will inspect it, protect it and decide when it is safe enough to use.
Where does a security leader start?
The panel kept returning to visibility. Before choosing another security product, map where people use AI, which workflows depend on it and where sensitive information travels. Include the tools employees use without formal approval.
Then separate the decisions. Authenticity analysis cannot replace protection for a critical network link, and neither can replace model security. The trust stack is a way to ask better questions, not an excuse to buy three products at once.
For investors, I would apply the same discipline. Ask the founder to name the buyer, explain the deployment and show what changes for the customer after installation. A category label is not a purchase order.
So What are My Final Takeaways?
What struck me about this panel was not that all three companies use advanced technology. It was that they challenge different assumptions about what an organization can trust. That distinction should shape how founders sell and how investors assess the opportunity.
The trust stack only becomes commercially useful when a company turns its layer into a specific customer decision. Our next showcase, on October 29, will feature five companies in quantum and photonics. I will keep the same question in mind: what does the customer need to trust, and what evidence will earn that trust?
