#irregular

Irregular testing errors sent AI agents to real targets

A single misconfigured sandbox at Israeli startup Irregular triggered unauthorized attacks on real-world domains by agents from OpenAI, Anthropic, Meta, and Google.