To address the ever-growing rail transport demand, the concept of Virtual Coupling train operations is gradually gaining ground within the railway industry. Thanks to a Vehicle-to-Vehicle communication, trains could be separated by less than an absolute braking distance and even form connected platoons to increase capacity at bottlenecks. However, a major concern about this concept regards the risk of safety violations in case of operational hazards pertaining to delays in train communication and control or emergency train stops. In this paper, the notion of dynamic safety margin is introduced for Virtual Coupling to dynamically adjust train separations so to always keep required safety distances also when hazardous operational events occur. The dynamic safety margin is embedded in a multi-state train-following model to analyse Virtual Coupling operations in presence of operational risk factors. A three-step methodology is applied in a real case study to fine-tune and verify the model, perform a sensitivity analysis, and identify capacity gains in several test scenarios including nominal and degraded traffic conditions. Results show that the use of a dynamic safety margin provides substantial capacity benefits to Virtual Coupling while respecting safe train distances even in case of sudden failures of the train control or communication systems. The notion of a dynamic safety margin can hence contribute to a safer version of Virtual Coupling operations and be considered by the railway industry in defining system requirements.