Gemini Smart Home Agent Ordered Cupcakes After Misreading a Dentist Appointment

A Gemini-powered smart home agent misinterpreted a dentist appointment calendar entry as a commerce instruction and placed an actual cupcake order, illustrating a real-world prompt injection failure. The incident highlights how LLM-based agents process structured inputs like calendar events as potential action triggers, blurring the line between data and instructions. Unlike enterprise agent platforms from vendors such as Zenity and Lakera, consumer smart home systems lack runtime policy enforcement, tool-call audit logs, or kill switches to catch such errors. This means users only discover a misinterpretation after the agent has already acted, with no visibility into what triggered the response. While the cupcake order caused no serious harm, security researchers warn the same vulnerability pattern could lead to unauthorized financial transactions or physical security breaches.