Irregular

From Pattern Labs to Irregular

Irregular, originally called Pattern Labs, was founded in 2023 by Dan Lahav and Omer Nevo, and came out of stealth in September 2025. According to the company's own official blog, its mission is "protecting the world in the time of increasingly capable and sophisticated AI systems."

What the company actually tests

Most model tests check answers to questions. Irregular instead builds simulated network environments and measures how far a model gets on the attack: whether it finds a weakness, exploits it, moves from computer to computer and covers its tracks. Model labs use results like these to decide which capabilities are too dangerous for open release.

$80 million raised, according to the company's official blog

According to the official blog it published itself, Irregular raised $80 million led by two funds, Sequoia and Redpoint, with the company already bringing in "millions in annual revenue." Calcalist's report on the same round named Sequoia as the lead, with participation from Wiz founder Assaf Rappaport and other investors. It is the official blog that adds Redpoint as an additional co-lead of the round.

Professional recognition: from OpenAI to the UK government

According to the same blog, OpenAI cites Irregular's evaluations in the system cards of o3, o4-mini and GPT-5; Google DeepMind researchers cited it in a paper on emerging AI cyberattack capabilities and used its platform; it co-authored a foundational paper with RAND on securing model weights; and it also works as a partner to the UK government in vetting cyber capabilities in frontier models.

2026: when the company itself became part of the story

During 2026 both OpenAI and Anthropic reported incidents in which models being tested in Irregular's test environment managed to get out of it onto the open internet, because of a configuration error in the environment's isolation. The incidents caused no proven harm, but they showed the paradox at the heart of the field: to know what an autonomous agent is capable of, you have to let it act in a realistic environment, and the more realistic the environment, the harder it is to make sure the agent stays inside it.

The Israeli context: a small market, a big impact

Most Israeli AI companies build a product for end customers, while Irregular sells an evaluation service to the handful of organizations in the world that develop frontier models. It is a small market in number of customers but significant in impact, since its results affect what is released to the public. Alice (formerly ActiveFence), headquartered in New York and Tel Aviv, also works with frontier labs, simulating malicious prompts and agent tasks for them to expose unexpected model behavior.