As assistive and collaborative robots become more ubiquitous in the real-world, we need to develop interfaces and controllers that are safe for users to build trust and encourage adoption. In this Blue Sky paper, we discuss the need for co-evolving task and user-specific safety controllers that can accommodate people's safety preferences. We argue that while most adaptive controllers focus on behavioral adaptation, safety adaptation is also a major consideration for building trust in collaborative systems. Furthermore, we highlight the need for adaptation over time, to account for user's changes in preferences as experience and trust builds. We provide a general formulation for what these interfaces should look like and what features are necessary for making them feasible and successful. In this formulation, users provide demonstrations and labelled safety ratings from which a safety value function is learned. These value functions can be updated by updating the safety labels on demonstrations to learn an updated function. We discuss how this can be implemented at a high-level, as well as some promising approaches and techniques for enabling this.