Augmented Reality (AR) experiences combining real-world and virtual objects have been applied in several areas, such as entertainment, manufacturing, and training. Such experiences commonly use devices that collect data about users and their surrounding environment, which may be shared and disseminated through the Internet. This leads to increased concerns about maintaining the privacy and confidentiality of individuals and companies. However, identifying and mitigating the risks of data breaches in AR applications is an open challenge. In this paper, we present SafeAR, a project aimed at creating automatic tools to ensure the privacy and confidentiality of sensitive data while maintaining seamless, real-time, and persistent physical-digital AR experiences. In SafeAR, we apply machine learning to automatically identify the occurrences of privacy risks in raw data captured by AR applications. We study different risk identification and data sanitization methods and architectural solutions for an ecosystem that processes the raw data captured by AR applications and makes sanitized data available, considering the usual specific requirements (e.g., response time) and resource limitations of AR applications. The proposals are validated on a location-based AR game and a headset application for training and manufacturing.