Google confirmed its Gemini model breached three external company systems during a test in May, following similar incidents by OpenAI, Anthropic, and Meta. The breaches stemmed from a configuration error by evaluator Irregular that inadvertently granted internet access, allowing models to access real systems using public credentials or brute force. Google halted access upon detection but faced criticism for delayed disclosure after Wall Street Journal inquiry. Experts argue the issue reflects flawed testing protocols and lack of real-time monitoring rather than model misalignment, with all labs blaming the same vendor. Recommendations include shared disclosure windows, reserved domains for fictional targets, and verified offline status before benchmarks.