---
title: OpenAI, Anthropic and Meta All Send Their Models to the Same Security Lab
description: Irregular tests frontier AI models for OpenAI, Anthropic and Meta. It ran the evaluations behind two of this month's breaches.
author: Darie Nani (Editor-in-Chief)
date: 2026-08-10T07:00:00.000Z
updated: 2026-08-11T17:04:44.796Z
canonical: https://www.sovereignmagazine.com/article/irregular-frontier-ai-security-lab-openai-anthropic-meta
image: https://cdn.nanimediahouse.com/pexels-a-man-in-sunglasses-intently-studies-a-vibrant-blue-holograp-5866051.jpg
categories: Artificial Intelligence
content_type: News
region: Global
publication: Sovereign Magazine
access: members
schema_type: Article
---

OpenAI, Anthropic and [Meta](https://www.sovereignmagazine.com/article/meta-muse-glimmer-open-30b-apache-2) have all pointed to the same outside firm after their AI models broke out during security testing this month: Irregular, a small Israeli lab that describes itself as the first frontier security lab. Over roughly two weeks each of the three companies disclosed that one of its models had reached the open internet during Irregular's evaluations, and Irregular says all three came from the same testing-environment fault, not a sophisticated escape from its sandbox.

Irregular, formerly called Pattern Labs, says it runs controlled simulations on frontier AI models, testing both their potential for misuse in cyber operations and their resilience when they are attacked. OpenAI cites its evaluations in the published system cards for GPT o3, o4-mini and GPT-5. It has co-authored a white paper with Anthropic on confidential inference systems and a paper with RAND on securing AI model weights against theft, Google DeepMind researchers have cited it in work on emerging AI cyberattack capabilities, and it advises government institutions including the UK government on vetting the cyber capabilities of frontier models. It has raised $80 million led by Sequoia and Redpoint and says it already makes millions of dollars a year.

## The Lab Assessed the Meta Model a Month Before the Breach

---

*This article is only available to registered readers. Visit the article URL to read the full piece.*