A series of incidents has kept the AI industry on edge for weeks, with AI models from Anthropic, OpenAI, Meta and Google attacking real companies during security tests—guessing passwords, harvesting credentials, breaking into production databases and uploading a malicious package to a public software registry. Many trails lead to Irregular, a Tel Aviv startup calling itself the world’s “first frontier security lab,” formerly known as Pattern Labs, which tests AI models for offensive cyber capabilities before release by building simulated environments and whose customers include Google, OpenAI, Anthropic and Meta. In several test environments, access to the real internet was open even though models’ instructions said they had none, so models mistook real companies for part of the simulation and attacked them. A Wall Street Journal overview lists eleven cases in which AI agents recently broke out of test environments; four trace back to Irregular. Anthropic disclosed three incidents across six test runs, the earliest dating back to April, plus a later fourth involving an older checkpoint of Claude Opus 4.6. OpenAI reported models in an Irregular environment mistakenly connected to the internet exploited a real website. Meta confirmed one of its models hacked an outside company during an external test. As early as May, Gemini hacked three companies during an Irregular “capture the flag” exercise.