Verax AI

**Product and the concrete problem it solves.** Verax AI sits in the fast-forming category of *GenAI security and assurance* — the control layer that enterprises need before they can safely put large language models (LLMs) into production. The concrete problem is that LLMs are non-deterministic and cannot themselves distinguish a verified fact from a confident fabrication, so as organizations move generative AI from sandbox pilots into live, customer- and employee-facing workflows they inherit a new class of operational risk: hallucinated or factually wrong answers, data leakage, prompt-injection and jailbreak attempts, biased or off-policy responses, and "shadow AI" — employees quietly piping sensitive corporate data into unsanctioned third-party models. Traditional MLOps and model-development tooling optimizes models *before* deployment; it does little once a model is live and behaving unpredictably in the wild. Verax's pitch is a purpose-built layer that closes that gap. Its platform — described across the company's materials as the **Verax Control Center** and the security-hardened **Verax Protect** — provides organization-wide discovery of where and how GenAI is actually being used, real-time monitoring of model behavior in production, and automated correction or blocking of unwanted outputs, so that CISOs and risk owners can say "yes" to AI adoption without accepting uncontrolled exposure.

**Core technology and how it actually works.** Architecturally, Verax deploys as an intermediary control plane *inside the customer's environment*, positioned between users/applications and the LLMs they call. Rather than retraining or fine-tuning models, the system observes traffic and enforces policy at runtime, which lets it work across heterogeneous model estates. Two design choices matter most. First, the platform is described as **LLM-agnostic and requiring no pre-training or lengthy configuration** — a deliberate bet that enterprises will run many models (commercial APIs, open-weight models, and internally hosted ones) and need one horizontal governance layer rather than per-model bespoke guardrails. Second, it emphasizes **lineage and traceability**: when a response is generated, Verax traces the relationship between the AI's output and its source data so that deviations, hallucinations, or unsafe content can be flagged, diagnosed, and — per the company — auto-corrected in real time without human intervention or slow manual tuning. On the security side, Verax Protect adds **shadow-AI discovery, prevention of unauthorized model training on corporate data, runtime protection, and identity-aware access and usage control** — i.e., tying every GenAI action back to who is doing it and enforcing organizational policy on that basis. The honest technical caveat is that runtime LLM guardrailing is a young discipline; the effectiveness of automated hallucination detection and auto-correction is workload-dependent and not something the public record independently benchmarks.

**Market, customers, and go-to-market.** Verax targets the enterprise buyer — specifically CISOs, heads of AI/data governance, and risk and compliance owners in organizations that want to deploy GenAI "in a secure and compliant fashion." The go-to-market is classic top-down enterprise security software: land with security and governance teams who are the internal gatekeepers for AI rollout, then expand as GenAI usage proliferates across business units. Notably, the founders report having validated demand by interviewing CISOs across diverse industries *before* building the product, which is the right sequencing for an enterprise-security company. The addressable market is large and expanding quickly — every regulated and data-sensitive enterprise adopting GenAI is a candidate — and it is being pulled forward by regulatory pressure (the EU AI Act, sectoral compliance regimes, and internal audit requirements) that makes an auditable AI control layer a near-necessity. The calibrated counterpoint: the public record does not disclose named production customers, revenue, or logo references, so commercial traction beyond the seed round should be treated as unproven, and the category is being contested by dozens of well-funded entrants simultaneously.

**Traction, funding, and third-party validation.** Verax was **founded in 2023** and announced a **$7.6 million seed round on October 30, 2024, led by TQ Ventures**, with participation from Concept Ventures, Cardumen Capital, Seedcamp, XTX Ventures, and **InMotion Ventures** (the corporate venture arm associated with Jaguar Land Rover). The syndicate is a meaningful signal: it blends specialist seed investors with a strategic corporate backer, and InMotion published a public thesis explaining that it invested because the market for *practical, commercial* LLM-assurance solutions "remains sparse" despite heavy research interest — an investor articulating genuine white space rather than a crowded me-too bet. The company launched its enterprise platform publicly in late October 2024. Beyond the seed round and platform launch, third-party validation is still early-stage: there are no publicly disclosed enterprise contracts, certifications (SOC 2, ISO, FedRAMP), or defense/government engagements in the record, and headcount is not publicly confirmed. This is a seed-stage company whose validation to date is investor conviction and founder pedigree rather than disclosed revenue or marquee customers.

**Founders and team background.** The strongest asset in the Verax story is its team. Co-founder and CEO **Leonid ("Leo") Feinberg** is a third-time founder with a real exit track record: he co-founded **CloudEndure**, where he served as VP of Product, and CloudEndure was **acquired by Amazon Web Services for a reported ~$250 million**; an earlier venture produced an exit to Limelight (~$25M). Co-founder and CTO **Oren Gev** worked alongside Feinberg at CloudEndure and subsequently at AWS before the pair left in 2023 to build Verax. That background — enterprise infrastructure and cloud-scale software, a prior category-defining product, deep AWS operating experience, and demonstrated enterprise-sales muscle — is precisely the profile that de-risks an early enterprise-security company. The team operates with an Israel R&D core (Tel Aviv) and a U.S. commercial presence (reported across Dallas and New York, with additional London footprint in some accounts), the standard Israeli-founded, U.S.-facing enterprise-software structure. The principal open question is scaling a go-to-market and research organization fast enough to keep pace with a category expanding — and consolidating — in real time.

**Competitive dynamics.** Verax competes in one of the most crowded frontiers in enterprise software, and its differentiation rests on being a horizontal, runtime, LLM-agnostic control layer rather than a point tool. (1) Against **AI-security specialists** such as Protect AI (acquired by Palo Alto Networks), Lakera, CalypsoAI, WitnessAI, and Israel's own Prompt Security, Verax competes on breadth (discovery + monitoring + auto-correction + access control in one plane) versus depth in a single threat like prompt injection. (2) Against **AI-governance/compliance platforms** (Credo AI, Holistic AI, Robust Intelligence — the latter acquired by Cisco), it competes on being a live runtime enforcement layer rather than a documentation/assessment tool. (3) The most serious structural threat is **platform absorption**: hyperscalers and model providers are shipping native guardrails (Azure AI Content Safety, AWS Bedrock Guardrails, Google/OpenAI safety tooling), and security incumbents are acquiring their way into AI security, so a horizontal startup must prove it is materially better and genuinely model-agnostic to avoid being commoditized by "good-enough" native controls. Verax's plausible edges are its no-pre-training/LLM-agnostic deployment, its lineage-based hallucination tracing, and a founding team that has built and sold enterprise infrastructure before.

**Defense, security, and resilience dual-use relevance.** Verax's dual-use relevance should be read as *real on the security/resilience axis and adjacency-grade on the defense axis.* The technology is, at its core, security software: it governs how sensitive organizations adopt a powerful and unpredictable new capability, prevents proprietary data from leaking into external models, discovers unsanctioned AI usage, and enforces identity-based policy and auditability. Those functions map directly onto the requirements that defense, intelligence, and critical-infrastructure operators will impose before they can field LLMs in mission contexts — trustworthy-AI assurance, hallucination and data-exfiltration control, red-team/jailbreak resistance, provenance, and policy enforcement are exactly the guardrails that make generative AI usable in high-stakes, low-error-tolerance environments. As allied governments push "responsible/assured AI" mandates, an independent, model-agnostic control plane is strategically relevant national-security infrastructure in the abstract. The calibration that keeps this honest: Verax has *no publicly disclosed defense or government customers, no fielded classified deployments, and no security certifications on the public record*. Its dual-use case today is the enterprise-security-and-resilience thesis, with defense as a credible future adjacency rather than a demonstrated capability.

**Growth stage, trajectory, and key diligence risks.** Verax reads as an **early-stage** company: founded 2023, one disclosed $7.6M seed (October 2024), a launched platform, a strong syndicate, and an exceptional founding team, but without disclosed revenue, named customers, certifications, or a Series A on the public record. The trajectory is promising precisely because the founders have run this playbook before and are early to a category that regulators and enterprise risk owners are forcing into existence. The key diligence risks are, in order: (1) **category commoditization** — hyperscaler-native guardrails and security-incumbent acquisitions could compress the standalone opportunity; (2) **crowding and consolidation** — dozens of AI-security startups are chasing the same CISO budget, and shakeout is likely; (3) **unproven traction** — commercial validation beyond investor conviction is not public, so revenue, retention, and logos remain unverified; (4) **efficacy risk** — automated hallucination detection and real-time auto-correction are hard technical claims whose production performance is not independently benchmarked; and (5) **positioning risk** — the "Control Center" vs. "Protect" framing suggests an evolving product identity that will need to settle into a durable wedge. Progression from here would be evidenced by a Series A, disclosed enterprise customers (ideally in regulated or public-sector verticals), security certifications, and independent evidence that the runtime guardrails materially outperform native alternatives.