What Irregular does

Irregular is a frontier AI security lab headquartered in Tel Aviv, Israel, founded in 2023 with a mission to evaluate and defend against offensive cyber capabilities emerging from advanced AI systems. The company builds proprietary evaluation frameworks—CyScenarioBench (scenario-based, multi-stage attack-tree evaluations), FrontierCyber (real-system offensive benchmarks with no planted vulnerabilities), and the Atomic Tasks Suite (network security, vulnerability research/exploitation, and evasion assessments)—that measure how well AI models can plan and execute realistic cyber operations. Irregular also publishes technical research on emergent agentic cyber behavior, password predictability, refusal policies, and AI-enabled vulnerability discovery; the company discovered three real zero-day vulnerabilities (in Soft Serve and QuickJS) during evaluation work.

The company's customers are concentrated among frontier AI labs (OpenAI, Anthropic, Google, Meta) and government research bodies (AISI), with joint research collaborations with partners such as Wiz. Irregular generates revenue through two primary streams: (1) SaaS subscriptions delivered via API for automated model evaluation, and (2) professional services engagements for collaborative research and custom evaluations with AI developers. Distribution is purely direct enterprise sales with no channel partners; marketing is primarily earned media and research-driven content. The firm operates with a small (11–50 employees), talent-dense team including AI scientists, senior R&D leaders, world debate champions, and senior cyber and national security experts.

Irregular closed an $80 million funding round in September 2025, co-led by Sequoia and Redpoint (with Swish Ventures participating), to scale the first dedicated frontier AI security lab. The firm was ranked #1 in Calcalist's 50 Most Promising Israeli Startups 2026, recognized on Forbes' AI 50 Brink List, and won Fast Company's World Changing Ideas 2026 award. As of mid-2026 the firm was actively publishing 1–2 research items per month, launching new benchmark products, and expanding collaborations across the frontier-AI ecosystem.

Name

Irregular

Legal name

Irregular

Website

https://irregular.cv

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

11–50 employees

Short description

Frontier AI security lab that evaluates AI models' offensive cybersecurity capabilities through proprietary benchmarks (CyScenarioBench, FrontierCyber, Atomic Tasks). Serves frontier AI labs (OpenAI, Meta, Anthropic, Google), governments, and enterprises via SaaS subscriptions and custom research collaborations.

Ownership category

akta.pro rank

Product category

AI Security / Cyber Capability Evaluation

NAICS

Security Systems Services (except Locksmiths) (561621)

SIC

Services-Computer Programming Services (7371)

akta.pro primary industry

Safety & Alignment Evaluation (red-teaming, harmful capability testing) (HDAAAMAL)

akta.pro secondary industry

Security Incident Response (IR) & Digital Forensics Services (BPAEADAI)

Keywords

• AI security evaluation
• Offensive cyber benchmarking
• LLM cyber capabilities
• AI red teaming
• Vulnerability research automation

Headquarters

HQ city

Tel Aviv

HQ country

Israel

HQ region

Middle East

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Operations

Revenue model

1. SaaS Subscription: Customers access products and services through SaaS subscription agreements with API-based data handling. Subscriptions include agreements detailing data collected and processing methods.
2. Research Collaboration & Evaluation Services: Collaborative research engagements with AI labs (OpenAI, Meta, Anthropic) for evaluating frontier models against proprietary security benchmarks. Work-for-hire evaluations and security research partnerships.

Pricing tiers

Go-to-market motion2 records

Distribution channels2 records

Marketing channels4 records

Core offering

Irregular operates as a frontier AI security lab that builds proprietary evaluation platforms (CyScenarioBench, FrontierCyber, and the Atomic Tasks Suite) to measure and monitor the offensive cybersecurity capabilities of frontier AI models. It delivers these capabilities to AI developers and enterprises through SaaS subscriptions accessed via API, alongside direct research collaborations and model evaluation engagements.

Product overview

Irregular operates as the first frontier AI security lab, offering a portfolio of proprietary evaluation platforms and research publications. The core offerings include CyScenarioBench (scenario-based multi-stage cyber evaluations), FrontierCyber (real-system offensive evaluations), and the Atomic Tasks Suite (task-level capability assessments). These platforms work together to provide comprehensive evaluation of AI models' cybersecurity capabilities, from isolated technical tasks to full attack campaign simulations. The company also publishes research on AI security threats and capabilities. Irregular's SaaS subscription provides API-based access to customers for data handling through evaluation subscriptions.

Differentiator

Problem solved

Functional benefit

Products and services

• CyScenarioBench Scenario-based evaluation framework that measures an LLM's ability to plan and execute multi-stage cyber scenarios under realistic constraints, evaluating cyber orchestration, branching-decision accuracy, constraint adherence, and recovery from state inconsistencies.
• FrontierCyber Benchmark that turns real-world systems (mobile devices, hosted software services, databases, networks, and routers) into offensive cyber capability evaluations for AI models, with verifiable objectives and standardized conditions.
• Atomic Tasks Suite Proprietary evaluation suite covering three domains: network security (standard attack sequences, network mapping, protocols, firewalls, file servers); vulnerability research and exploitation (reverse engineering, code scrutiny, cryptography, exploit development); and evasion (detecting a model's ability to avoid security monitoring).

Quantifiable outcome

• AI agents solve 9 of 10 web security challenges at under $1 per success vs ~$100,000 for human researchers
• +2 more outcomes

Named customers6 records

Segments2 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability6 records

Feature5 records

Partnerships

Five partnerships are on record, tiered flagship and core.

• MetaflagshipStrategic or Co-development Partner · 15 April 2026Collaborative testing of Meta's Muse Spark model against Irregular's offensive security evaluation suites (CyScenarioBench and Atomic Tasks). Testing conducted as part of Meta's Advanced AI Scaling Framework.
• WizcoreStrategic or Co-development Partner · 29 January 2026Joint research collaboration testing AI agents (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro) on web security CTF challenges. Study found AI agents solved 9 of 10 challenges at low cost, demonstrating AI-enabled offensive security capabilities.
• AISI (AI Safety Institute)coreStrategic or Co-development PartnerJoint work on evidence that standard evaluation budgets underestimate frontier model cyber capabilities. Collaboration on presenting findings about inference scaling in AI cyber tasks.
• OpenAIflagshipStrategic or Co-development PartnerWorked with OpenAI to evaluate GPT-5.5 using Irregular's internal offensive security benchmarks, assessing model's cyber offensive capabilities.
• AnthropicflagshipStrategic or Co-development PartnerOngoing collaboration on model evaluation testing Claude family models against proprietary security benchmarks.

Scale indicators7 records

Recent moves7 records

Expansion highlights5 records

Direct peers

• HiddenLayer: AI security company offering adversarial testing, model vulnerability scanning, and threat detection for machine learning systems. Direct commercial peer in the AI security evaluation category, focused on enterprise defensive use cases.
• METR (Model Evaluation and Threat Research): Non-profit research organization that builds rigorous task-based evaluations of frontier AI capabilities, including autonomous task horizons used by OpenAI and Anthropic. Most direct comparable to Irregular on the methodology side, though operating as a non-profit rather than a commercial vendor.
• Apollo Research: AI safety evaluations lab focused on scheming, deception, and high-risk capabilities in frontier models. Closely comparable mission to Irregular, with similar government and frontier-lab customer base and published reports on dangerous capabilities.
• Redwood Research: AI safety research organization that studies and mitigates dangerous capabilities in frontier models. Operates in the same category of work as Irregular — evaluations and applied safety research for frontier AI developers.
• CalypsoAI: AI security and evaluation platform focused on testing, validating, and securing generative AI systems for enterprises. Overlaps with Irregular on AI model evaluation and red-teaming, but targets enterprise security buyers rather than frontier labs and regulators.
• Robust Intelligence: AI security and robustness platform (acquired by Cisco) that tests and protects AI applications from adversarial attacks and model vulnerabilities. Comparable in scope to Irregular's model-evaluation work but historically more focused on production AI applications.

Emerging players

• Lakera AI: AI security company offering Gandalf, a prompt-injection and LLM red-teaming platform. Comparable on the adversarial-evaluation axis but more narrowly scoped to generative AI application security rather than full-spectrum cyber capability assessment.
• Dreadnode: Startup focused on AI/ML red-teaming and offensive security tooling. Smaller and earlier-stage than Irregular, with overlapping interest in evaluating AI for offensive cyber capabilities.

Broad incumbents

• Wiz: Cloud security platform that has co-published research with Irregular on AI agents and web security. Comparable customer segment (cloud and security-conscious enterprises) and adjacent scope, but broader product surface than AI-specific evaluation.
• HackerOne: Established bug bounty and offensive security testing platform. Touches a similar offensive-security theme and customer base as Irregular, but covers human-driven vulnerability disclosure rather than AI capability evaluation.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights8 records

Customer concentration

• Xx.com/irregular
• LinkedInlinkedin.com/company/irregular

Revenue estimate

Valuation estimate

Number of profiles

Profiles1 record

Funding overview

Funding rounds1 record

Investors3 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Irregular