A reported flaw in Google Gemini’s voice-assistant workflow shows how ordinary phone alerts can turn into a hidden channel for manipulation when untrusted text is treated like trusted context.