OMER NEVO

Behavioral Archetype

THE MAN WHO WIRED THE CAGES — Subject is the technical co-founder of the company whose simulated networks sit inside the pre-release cyber evaluations of the largest labs. His co-founder, Dan Lahav, is the company’s chief executive and its voice to governments; Subject is its chief technology officer and the one who describes, in his own words, what the ranges do. In 2026 models reached the real internet from inside evaluations built on that platform, and the company published its postmortem unsigned. The record names the company, not the engineer. This profile scores the position, not the man.

Essence Indicators

• Co-founder and CTO, Irregular. Globes, on the 2025 funding round: “founded by CEO Dan Lahav and CTO Omer Nevo” (Globes, Sept 17 2025). Subject’s own byline carries the same title: “The author is co-founder and CTO of Irregular.” (The Sidechannel, opinion piece listed by Irregular as Feb 16, 2026). The Tel Aviv AI Security Forum, December 10, 2025, lists him as “Co-Founder and CTO @ Irregular” (AI Security Forum).
• Founded in 2023, as Pattern Labs. “Formerly known as Pattern Labs, Irregular is already a significant player in AI evaluations” (TechCrunch, Sept 17 2025). Calcalist dates the founding to 2023; Bizportal, in a Hebrew-language piece, says the company “was founded about two years ago by Dan Lahav and Omer Nevo” (Calcalist; Bizportal). Irregular’s own About page names no founders and gives no founding date (Irregular).
• Before Irregular. “Nevo is a serial entrepreneur who served as a development manager at Google Research” (Calcalist Tech, Sept 17 2025); Calcalist’s April 2026 profile adds that he founded NeoWize, “sold to the cosmetics company Il Makiage” (Hebrew original in the Sources line), then joined Google, where he led a research team on wildfires (Calcalist, April 21 2026). These are press accounts. No Irregular page or post of his own states them.
• Debating. “Nevo is a world debate champion” (Calcalist Tech). The same pages say he and Lahav met through competitive debate.
• The money. $80 million “in a round led by Sequoia Capital” and Redpoint Ventures (TechCrunch). Full detail on the Irregular profile.
• The outputs. Irregular’s ranges and tests appear in the labs’ own system cards: GPT-5.5, GPT-5.6 and GPT-6 Astra (OpenAI) name it as an outside cyber evaluator, and Google’s Gemini 4 Argon report says CyScenarioBench was co-developed with Irregular (The Evaluator Alliances).
• The leaks. In the company’s August 14, 2026 account, a fictional target name “unintentionally coincided with a real domain.” The post is unsigned and names neither founder (Irregular). Full record: AI Agent Incidents.

Social Persona / Impression Management

Immediate impression: The engineer who explains. Where the CEO frames the company’s mission, Subject describes the mechanism: simulated networks, attacker and defender agents, and what happens when a new model is dropped in.

Energy: Plain, concrete, defender-first. He quotes no threat actor and names no lab in the statements on file; the unit of his argument is “the defenses.”

Impression management strategy: Credibility by specificity. “We have complex network simulations where we have AI both taking the role of attacker and defender,” he told TechCrunch, “So when a new model comes out, we can see where the defenses hold up and where they don’t.” The register is a test engineer’s, which is also the position a contractor inside a release decision needs to occupy: the one who reports, not the one who decides (TechCrunch).

Forensic Archetype Comparison

Psychometric Assessment

Scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN):

Dark Triad:

MBTI: INTJ (“The Architect”) — builds the model of the system first (attacker agents, defender agents, a simulated company) and reports what it does. The position rewards exactly that, and a test engineer is selected for it.

Threat Assessment

Score rationale: set below Dan Lahav (7/6/7, hazard 64), who holds the same company’s government, EU and press seat, and in line with Matt Fredrikson (7/6/6, hazard 60) at Gray Swan, the company’s nearest peer. Attack is 6 rather than 7: the record shows him explaining the ranges, not setting policy on them. Hp is 7 because the platform is the one thing the company cannot lose.

Alignment Analysis

Stated alignment: Defenders should be using AI as fast as attackers are. In Subject’s words: “Are we helping defenders utilize AI fast enough to be able to keep up with what offensive actors are already doing?” And of the field’s pace: “Even for a lot of seasoned professionals who have seen both AI and cybersecurity, it has been genuinely surprising” (Fortune, Jan 29 2026). On capability: “We’re seeing more and more that models are able to solve challenges that are genuine expert level” (same). His 2026 opinion piece argues the near-term beneficiary is the bottom of the attacker pool: “Frontier AI is empowering the least capable cyber actors,” a group it describes as “opportunistic criminals: large in number, modest in sophistication” (The Sidechannel).

Observed alignment: The company he co-founded is paid, in the labs’ own documentation, to run cyber evaluations of their models, and has declared a new priority for the field in its own words: “We believe that the community urgently needs to discuss internet access standards during pre-deployment evaluations” (Irregular). Its own account of cause: “Due to human oversight, however, it unintentionally coincided with a real domain,” and “Ultimately, most of the issues we’ve discovered were due to internet access controls” (same).

Gap assessment: The record cannot settle, and does not need to, whether the defender-first argument is conviction or business development; it works as both. The gap that matters is structural and belongs to the company: an evaluator is most useful when its test is contained, and the test that measures whether a model can breach a network is the test whose containment failed. That the company reported it, fixed its review process and called for standards is documented; that any named individual designed or approved the environments is not, and is not asserted.

Interest: each position, with what adoption would pay the company he co-founded, then the best case.

1. “Defenders must adopt AI fast enough.” Adoption pays any vendor of AI-assisted defence and any evaluator whose scenarios train or test defenders; Irregular’s own research describes AI in “the role of attacker and defender” in its simulations (TechCrunch). His best case: attackers are already using the tools, defenders institutionally move slower, and he says so as the person measuring expert-level performance (Fortune). Same-rule factions: the large security vendors that bought AI-security firms through 2025 (The AI Evaluators) gain from the same demand, as do the labs, whose dual-use capability is reframed as a defensive input.
2. Independent pre-deployment cyber evaluation. The company’s revenue model is that evaluation; the UK award, the funding round and the system-card citations all rest on the premise that labs need outside range-builders (Contracts Finder; TechCrunch). A rule requiring it, or a government institute buying it, pays suppliers of that service. His best case: a model’s offensive ability is unknowable without a realistic range, and a lab grading itself has the obvious conflict. Same-rule factions: the other contract evaluators, Gray Swan among them, hold awards in the same January 2024 set; the labs gain a citation that lets them say outsiders tested the model.
3. Standards for internet access in evaluations. The company asks the field to set them (Irregular). A standard fixes the specification for the ranges the company builds, and moves the question of containment from one firm’s practice to a field-wide norm. Its best case: the same post says internet access is “at times critical for realistic evaluations,” so no single evaluator should decide the tradeoff alone. Same-rule factions: labs, whose incidents become a standards question rather than a vendor question, and rival evaluators, who inherit the same constraint.

Alliance Map

Ties on several axes run at once. The Corporate role that makes him technical lead of the evaluator is the same company that holds a UK government award, takes capital from two of the largest venture firms, and appears in several competing labs’ documentation: the evaluator is paid by, funded alongside, and cited by the parties it grades. The one ratchet-network tie in his own record, a seven-year board role in an effective-altruist organization that ended in September 2026, sits beside the evaluator layer described in The AI Evaluators, whose funders overlap that milieu; no coordination is shown. The overlap, not any single row, is the finding.

Convergent Drive Classification

Self-preservation: The platform survived the incident record: the company fixed its review, added a dedicated team and kept its place in the labs’ documentation (Irregular). Goal preservation: Defenders keep pace with attackers; his Fortune line is the same sentence in 2026 as the company’s pitch in 2025. Resource acquisition: $80 million from Sequoia and Redpoint, and an evaluation award from a government institute (TechCrunch). Self-improvement: Environments checked for domain overlaps before each run, manual review of model actions expanded.

Subject is not an AI system. The drives appear anyway — in the engineer whose product is the cage, and whose cage is what the models were measured against.

Sources: Irregular raises $80 million — TechCrunch (Sept 17, 2025); Israeli AI security lab Irregular raises $80m — Globes; Calcalist Tech (Sept 17, 2025); Calcalist (April 21, 2026; Hebrew: “שירת 12 שנה ביחידה 8200”; “הקים את NeoWize שנמכרה לחברת האיפור איל מקיאג’”); Bizportal (Hebrew: “החברה נוסדה לפני כשנתיים על ידי דן להב ועומר נבו”); The Sidechannel — “Frontier AI is empowering the least capable cyber actors”; Fortune (Jan 29, 2026); Tel Aviv AI Security Forum 25; Addressing Recent Incidents — Irregular; Irregular news; UK DSIT award — Contracts Finder; LinkedIn (own page).