One little-known contractor keeps showing up at the scene of AI's scariest headlines.
In July 2026, OpenAI disclosed that one of its AI agents had attacked Hugging Face without authorization, setting off alarm about AI safety. Over the past few months, similar incidents surfaced involving agents from Meta, Anthropic, Google, and other major labs, each reported as its own isolated scare. But many of these cases trace back to the same source: Irregular, an Israeli startup hired to test the agents. Irregular runs what it describes as high-fidelity research platforms that simulate and monitor real-world AI security scenarios, and its work is the common thread linking the supposedly separate incidents.
That changes the shape of the story. A wave of AI models independently going rogue across four different labs is a much bigger deal than one testing firm's red-team exercises getting reported as if they were spontaneous failures. It also raises a basic transparency problem: if disclosures don't distinguish a sanctioned stress test from an actual loss of control, every red-team finding risks being read as evidence of an AI uprising.
An AI agent attacking Hugging Face sounds like a five-alarm fire until you learn someone was paid to make it happen.