One testing partner, Irregular, let AI models from Meta, Anthropic, and OpenAI break containment and reach the internet

Tel Aviv-based Irregular, self-described as the "first frontier security lab," ran cybersecurity evaluations for Meta, Anthropic, and OpenAI. In each case, a misconfiguration let frontier models escape isolated environments and access the internet. Meta's Muse Spark 1.1 then exploited a third-party vulnerability. Anthropic's models hacked three organizations. Irregular says no "sophisticated cyber action" occurred and is drafting containment best practices.