Problem

Irregular addresses the security gap created by increasingly capable frontier AI systems. As models become able to conduct multi-step cyber operations, discover and exploit vulnerabilities, evade defenses, and interact with real infrastructure, conventional security controls and pre-release testing can fail at multiple points. The central pain is high-consequence uncertainty: AI labs and operators need to know what a model can do under realistic attack conditions before releasing or deploying it. The killer use case is an independent, pre-deployment security evaluation that exposes cyber capabilities and emergent misuse risks early.

The economics are becoming more urgent because AI can make offensive security work cheap and scalable. In Irregular's testing, models solved nine of ten realistic web-security challenges, often at a cost below $1 or between $1 and $10, while the corresponding real-world bounty values ranged from thousands to tens of thousands of dollars. That asymmetry means even inconsistent model capability can create material security exposure when an attacker can run many attempts cheaply.

Product / Service

Irregular operates as a frontier-security research lab combined with an evaluation platform and specialized assessment service. Its platform can connect directly to a model or to the surrounding agent scaffolding, then run high-fidelity scenarios covering vulnerability discovery and exploitation, evasion of monitoring systems, reconnaissance, tool use, malware development, and situational awareness. Its environments use clear success conditions, such as capture-the-flag objectives, so evaluators can distinguish a genuine compromise from an agent's speculative or noisy report. Irregular also developed SOLVE, a scoring framework for measuring the difficulty of discovering and exploiting vulnerabilities.

The benefit is an objective, repeatable way for AI developers to measure both offensive capability and defensive robustness before deployment. Irregular can simulate networks in which AI systems act as attackers and defenders, assess a model across realistic threat scenarios, and provide independent evidence for safety evaluations and system cards. The commercial delivery model is therefore primarily lab-led B2B work—paid evaluations, research, and security frameworks for frontier-model developers and government or enterprise stakeholders—rather than a conventional standalone software product.

Market

Irregular competes in the emerging frontier AI security and AI-evaluation market, spanning model red teaming, cyber-capability testing, misuse evaluation, and deployment-safety research. Its closest overlaps are organizations such as METR, which evaluates broad autonomous capabilities and AI-enabled R&D, and Apollo Research, which focuses on pre-deployment testing for deception, evaluation awareness, and misaligned behavior. Irregular is more specifically differentiated around cybersecurity, offensive capabilities, and real-world security scenarios; adjacent AI red-team providers such as Gray Swan also compete for parts of the broader adversarial-testing budget.

Irregular is not pre-revenue. The company, formerly known as Pattern Labs and founded in 2023, announced $80 million of funding in September 2025 and said it had already reached millions of dollars in annual revenue. Its evaluations have been cited in OpenAI system cards, its SOLVE framework has been used by Anthropic and the UK government, and it has worked with OpenAI and Anthropic. By July 2026, Anthropic described Irregular as a third-party evaluation partner, while OpenAI's GPT-5.6 safety materials continued to identify it as a frontier AI security lab—evidence of meaningful adoption by leading model developers.