Abstract: Multi-robot systems demonstrate significant collaborative advantages in scenarios such as intelligent manufacturing, disaster rescue, and unmanned combat. However, as application environments shift from structured spaces to open, adversarial, and highly dynamic settings, system security has become a core bottleneck constraining their large-scale deployment. Based on threat sources and mechanisms, security challenges in multi-robot systems can be categorized into three dimensions: collision avoidance safety in physical space, fault-tolerant safety for system functionality, and cyber resilience in information space. From a unified cyber-physical integration perspective, we systematically review the evolution of safety control in multi-robot systems, covering technological progress from heuristic methods to formal control, from robust fault tolerance to distributed resilient control, and from model-driven to learning-enhanced approaches. The study focuses on key technical areas including safety-critical collision avoidance control, exemplified by formal methods such as control barrier functions and their data-driven enhancements; fault-tolerant control, covering fault detection and diagnosis, distributed cooperative fault tolerance, and system reconfiguration; resilient control, involving attack detection and identification, secure state estimation, and resilience recovery strategies; and safe reinforcement learning, which encompasses safety shielding, constrained optimization, and trust mechanisms. Through a systematic analysis of the theoretical foundations, representative advances, and applicability boundaries of these directions, the paper identifies core bottlenecks in current research, including safety–performance trade-offs, handling multi-source uncertainties, cross-layer cooperative design, and scalability. Finally, we discuss future research directions, such as the deep integration of learning with formal guarantees, cross-domain resilient architectures, and self-explainable safety decision-making, providing theoretical references and technical guidance for reliable autonomy in multi-robot systems.