Google Says Gemini Accessed Three Real Companies During AI Cyber Test

Google has confirmed that a Gemini model accessed systems at three real companies during a May cybersecurity evaluation, after a test designed for a fictional target reached the open internet. The episode did not reportedly cause damage, but it sharpens a central safety question for agentic AI: a model can follow the task it believes it has been given while the surrounding authority, scope and incentives are wrong.