When OpenAI and Anthropic each ran into significant safety incidents during model testing and deployment, the AI industry got a sharp reminder that building capable models and safely evaluating them are two very different engineering challenges. That gap is exactly where Irregular Expression has planted its flag. The startup, which specializes in automated red-teaming and adversarial testing for large language models, is now drawing serious attention from the labs that can least afford to get this wrong. According to a Calcalist report, Irregular has been working directly with leading frontier AI developers as demand for rigorous pre-deployment evaluation accelerates across the industry.
The timing is not incidental. As AI systems grow more powerful and are embedded in higher-stakes environments — customer service, healthcare triage, legal research — the cost of a failure mode discovered after deployment rather than before it has become existential for reputations and regulatory standing alike. The pressure is amplified by emerging government frameworks in the EU and the United States that increasingly require documented safety evaluations before certain AI systems can be commercially released. For context on how security-focused startups are capitalizing on AI’s expanding attack surface, Future Wire previously covered Oligo Security and its $60 million raise to tackle AI-assisted vulnerabilities at the runtime layer — a complementary but distinct problem from what Irregular targets at the pre-release evaluation stage.
What Irregular Actually Does — and Why It’s Hard
Red-teaming AI models manually is slow, expensive, and inconsistent. Human testers can probe a model for days and still miss entire categories of failure, particularly edge cases that emerge only at scale or under specific combinations of context and instruction. Irregular’s platform automates this process, generating adversarial prompts systematically across a broad taxonomy of risk categories — jailbreaks, harmful content elicitation, data leakage, and behavioral inconsistencies — and doing so at a speed and coverage depth that human teams cannot match. The Calcalist report describes the company as operating at the intersection of AI capability evaluation and cybersecurity methodology, a niche that barely existed three years ago and is now becoming a procurement line item for every serious AI lab.
The technical challenge is recursive in an almost elegant way: to test an AI system effectively, you need AI systems that are themselves capable of generating sophisticated, novel attack vectors. Irregular’s approach leans into this dynamic rather than fighting it. By training specialized models to probe for weaknesses in production-grade LLMs, the startup can surface failure modes that rule-based or template-driven testing frameworks would miss entirely. This positions Irregular not just as a compliance vendor but as an active research partner for labs trying to understand the behavioral envelope of their own models before anyone else does.
A Market Forming in Real Time
The competitive landscape for AI safety tooling is still taking shape, but it is filling in fast. Established cybersecurity players are beginning to extend their platforms toward model evaluation, and a wave of purpose-built startups — Irregular among the most visible — are moving to define the category before incumbents can absorb it. Enterprise security spending on AI-specific tooling is expected to grow substantially through 2026 as regulatory requirements harden and insurance underwriters start asking pointed questions about pre-deployment testing documentation. The same structural tailwind lifting agentic AI security companies — like those covered in Future Wire’s recent piece on $125 million raise — is running underneath Irregular, even if the two companies are solving different layers of the same problem.
Irregular has not publicly disclosed its funding figures or valuation, but the Calcalist reporting makes clear the company is actively engaged with top-tier frontier AI developers, which serves as its own signal of commercial traction. In a market where credibility is the scarcest resource, getting OpenAI- and Anthropic-adjacent work on your client list early is worth more than most seed rounds. The race to make AI safe enough to deploy at scale has no shortage of entrants, but the teams that can actually break sophisticated models on demand — reliably, repeatedly, and faster than the bad actors can — are going to be very hard to displace once they’re embedded in a lab’s development pipeline.
