One company is at the center of a wave of rogue AI attacks

In July, OpenAI said its AI agents had attacked Hugging Face without permission. That set off the usual wave of concern about AI safety, and then more reports followed: similar incidents tied to agents from Meta, Anthropic, Google, and others.

At first glance, those disclosures looked like a messy series of unrelated failures. But the detail that ties them together is stranger than that. A single company shows up again and again as the place where these agents were being tested.

That company is Irregular, an Israeli startup that stress-tests AI models in what it describes as high-fidelity research platforms. The pitch is to simulate and monitor real-world AI security scenarios. In other words, the kind of controlled environment where companies expect to catch bad behavior before it escapes into public view.

Instead, the testing itself has become part of the story. If the same shop is surfacing across separate AI incidents, then the issue may not be one bad model or one bad actor. It may be how little anyone really knows about what these agents do once they’re let off the leash, even briefly.