Dan Lahav

CEO of Irregular, the Tel Aviv firm that has become the default third-party cyber evaluator named in frontier-lab system cards. He sells expert-built evaluation scenarios, not labour hours, to the most competitive buyers in the market.

Background

Former IBM Research AI (Haifa) researcher; co-author of the Project Debater paper 'An autonomous debating system' (Nature, March 2021), which is the 'cover of Nature' credit in Sequoia's profile. Sequoia says he started working in tech at 14 and has a long-standing interest in superintelligence. In early investor meetings he demonstrated AI autonomously bypassing traditional endpoint security.

Co-founded Pattern Labs (later Irregular) with Omer Nevo in 2023.

What they run now

Keeping Irregular the trusted shared evaluator for competing labs after the July 2026 containment incident; expanding from benchmarks (SOLVE, CyScenarioBench) to evaluations on real systems (FrontierCyber) and to field-level agenda-setting with RAND.

Career

2023 – presentCo-founder & CEO, Irregular (Pattern Labs)
—Researcher (Project Debater), IBM Research AI (Haifa)

On the record

Offense-defense balance · 2026-08

AI is currently scaling offensive cyber capability faster than defensive capacity; argues for 'differential defensive cyber acceleration' now rather than waiting for an optimistic end-state. source

Mission relative to labs · 2025-09

If labs build ever more capable models, Irregular's goal is to secure them. source

Security stack · 2025-09

Expects human-on-AI and AI-on-AI interaction to drive much economic activity and to break the existing security stack. source

Interviews & talks

Irregular raises $80M to secure frontier AI models (interview) — TechCrunch, 2025-09

Connections

Omer Nevo — co-founder & CTO
Shaun Maguire, Dean Meyer (Sequoia) — lead investors
Assaf Rappaport (Wiz) — investor

Why it matters here

No overlap with creative data, but Irregular is a strong template: a small expert team whose benchmark became the referenced standard in labs' system cards, valued on that citation rather than on headcount. That is the 'own the evaluation standard' path Julian could copy for visual taste, where no such standard exists yet. The July 2026 incident also shows the operational risk of being the single shared vendor.

How to reach

Cold; Israel-based. Speculation: engages on evaluation methodology and field-building, as the RAND agenda and essays suggest; a question about how they got the first lab to cite a vendor benchmark is the useful ask.

What we could not establish

No verified X handle or LinkedIn URL.
Education and dates at IBM not verified.
Revenue undisclosed; Sequoia said only 'millions in revenue' in Sept 2025.