---
title: One Outside Lab Ran the Safety Tests in Both the Meta and Anthropic AI Breaches
description: Irregular ran the security evaluations behind both the Meta and Anthropic AI breaches. The lab says neither was a sandbox escape.
author: Darie Nani (Editor-in-Chief)
updated: 2026-08-09T23:13:40.409Z
canonical: https://www.sovereignmagazine.com/article/irregular-ai-safety-tests-meta-anthropic-breaches
image: https://cdn.nanimediahouse.com/irregular-ai-safety-tests-131935.webp
categories: Artificial Intelligence
content_type: News
region: Global
publication: Sovereign Magazine
access: members
schema_type: Article
---

Meta said on Wednesday that one of its AI models reached the open internet during a security test and broke into another organisation's systems. A week earlier, Anthropic said much the same thing had happened to one of its models, which got into three companies' systems. The evaluations in both cases were run by the same outside company, a small Israeli security lab called Irregular.

Irregular gives the same account of both incidents. It told Reuters that the Meta case did not involve "a sandbox escape or a sophisticated cyber action", and that it was "the exact same evaluation-environment issue that was already disclosed by Anthropic last week".

## The Model Involved Was Meta's Most Capable Coding System

---

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