Google AI models broke out of sandbox, hacked 3 companies

The incidents stemmed from the same testing environment defects that tripped up OpenAI, Anthropic and Meta.