Voice-based virtual personal assistants such as Amazon’s Alexa or Google Assistant have become highly popular and are used for diverse daily tasks ranging from querying on-line information, shopping, smart home control and a variety of enterprise application scenarios. Capabilities of virtual assistants can be enhanced with so-called Skills , i.e., programmatic extensions that allow thirdparty providers to integrate their services with the respective voice assistant.

In this paper, we show that specially crafted malicious Skills can use the seemingly limited Skill interaction model to cause harm. We present novel man-in-the-middle attacks against benign Skills and Virtual Assistant functionalities. Our attack uses loopholes in the Skill interface to redirect a victim’s voice input to a malicious Skill, thereby hijacking the conversation between Alexa and the victim. To the best of our knowledge this is the first man-in-the-middle attack targeting the Skill ecosystem. We present the design of our attack and demonstrate its feasibility based on a proof-of-concept implementation attacking the Alexa Skills of a smart lock as well as a home security system.