Found Zero-Days in Gemini. Then Raised $51.5M in 4 Months.

01 /The Concept

The penetration test has always had a structural problem: it happens once a year, covers the surface area a team of contractors could reach in two weeks, and produces a PDF report that is stale before engineering closes the most critical findings. The attack surface deployed since the last test is untested. The business logic flaw that requires understanding what the application is supposed to do before finding how it can be made to deviate from that intent has almost certainly never been looked at.

Novee was founded in May 2025 by Ido Geffen, Gon Chalamish, and Omer Ninburg – three former leaders in IDF offensive cyber operations. The product is a continuous AI penetration testing platform built on a proprietary model trained specifically for offensive security tasks, not a general-purpose LLM adapted for security use. The company benchmarked its model against Gemini 2.5 Pro and Claude 4 Sonnet on constrained web exploitation tasks and claims to outperform both by more than 55%, reaching approximately 90% accuracy against approximately 65% for frontier models on the same challenges.

The platform runs continuously across web applications, mobile apps, AI-native apps, APIs, and external attack surfaces. It chains exploits, reasons about application behavior to find business logic flaws that signature-based scanners miss by design, and returns findings with working proof-of-concept exploits and reproducible reproduction steps. Every fix gets automatically retested.

Enterprise controls are built for regulated customers: role-based access, full audit logs, reviewable test plans, rate-limited testing with no destructive payloads and no data exfiltration. Deployment spans SaaS, Bastion Node, and on-premises. Customer data never trains the model.

The founders published zero-day research in Gemini, Cursor, Microsoft, and Google's Python infrastructure, and presented at DefCon 2026. Four months after founding, Novee raised a $51.5M Series A led by YL Ventures and Canaan Partners, with Oren Zeev participating.

02 /Why It Matters

The four-months-to-$51.5M reflects a specific investor pattern rather than traction metrics alone. YL Ventures and Canaan Partners have backed Israeli offensive-background security teams at early stages before – Wiz, Island, and Perimeter 81 among the outcomes from that pattern. The IDF offensive cyber background, purpose-trained model, credibility built on published zero-day research, and DefCon presence map precisely onto the early profiles of those companies. The bet is on team and technical differentiation before revenue.

What makes the differentiation tractable: the gap between a purpose-trained offensive security model and a general model running security tasks is not a benchmark abstraction. General models reason about vulnerability classes learned from public sources – the known CVE categories, the common misconfigurations, the standard injection patterns. A model trained specifically on exploit construction, application behavior analysis, and vulnerability chaining develops coverage on the hardest class of findings: those that require reasoning about application intent before finding deviations from it.

Business logic flaws are the clearest example of that gap. A scanner can verify whether a JWT signature is checked. It cannot reason that a multi-tenant SaaS application has an authorization model that allows tenant A to access a resource owned by tenant B by abusing the structure of session identifiers. Finding that class of vulnerability requires understanding what the application was designed to do first. That's what the offensive background builds and what a model trained on those tasks can acquire in a way that signature-based tools structurally cannot.

The regulatory catalyst compounds the market timing. DORA, mandatory for EU financial entities since January 2025, requires threat-led penetration testing at specified intervals. PCI DSS 4.0, in full enforcement since March 2025, expanded its penetration testing requirements. NIS2 across the EU. SEC cybersecurity disclosure rules in the US. The compliance environment is pushing regulated enterprises from annual pentests toward continuous testing mandates – which converts Novee from a premium security capability into a compliance budget line.

03 /Opportunities

Continuous AI pentesting generates two categories of data that are separately valuable beyond the security findings themselves.

The first builds a market intelligence product. Aggregated, anonymized finding patterns across deployments – which vulnerability classes appear most in which technology stacks, which development practices correlate with lower finding rates, which frameworks produce the most business logic exposure – would be the most granular dataset on software security ever assembled from live production observation rather than breach statistics or survey data. Cyber insurers pricing risk, application security tool vendors validating claim coverage, and development platform builders measuring security impact of tooling choices would all pay for that signal.

The second is the compliance positioning play. SOC 2, ISO 27001, PCI DSS, DORA, and NIS2 all require penetration testing at intervals. A platform generating continuous evidence, mapping findings to specific control requirements, and producing audit-ready documentation on demand converts the compliance team from a reluctant secondary buyer into the primary budget owner – a position that survives security budget contractions in a way that discretionary tooling does not.

The entry angle that doesn't require competing with Novee directly: build the remediation intelligence layer between a continuous testing platform and engineering. Current pentest workflows produce findings; they don't produce the full context an engineer needs to fix them correctly without introducing a new vulnerability in the process. A system that takes a validated finding and generates fix specifications specific to the customer's codebase, framework, and deployment environment – with automated retesting built in – occupies a gap that every continuous pentesting platform, including Novee, has so far left unaddressed.