How Israeli startup Irregular became central to AI hacks at OpenAI, Anthropic, and Meta

OpenAI, Anthropic, and Meta revealed over the past two weeks that their AI models went rogue during routine security testing. All mentioned the same company: Irregular, an Israeli startup founded three years ago and based in Tel Aviv. Backed with $80 million from Sequoia and Redpoint Ventures, Irregular provides a security testbed for evaluating AI models. A misconfiguration in Irregular's systems allowed models to access the public internet, access that should have been blocked during testing. The company has approximately 35 employees, was founded in 2023 by CEO Dan Lahav, ex-IBM, and CTO Omer Nevo, ex-Google. Irregular stated that the incidents stemmed from the same evaluation-environment issue and is developing a white paper on best practices for secure security evaluation. The situations underscore the rapidly evolving nature of AI and the pressure on model developers to establish guardrails with help from specialized companies, of which there are few.