Google DeepMind revealed Gemini Robotics 2.5, a vision-language-action model that generates robot actions, and ASIMOV-Jailbreak, an open-source benchmark of prompts designed to trick robotic AI into unsafe actions