In large-scale, self-organized and distributed systems, such as peer-to-peer (P2P) overlays and wireless sensor networks (WSN), a small proportion of nodes are likely to be more critical to the system's reliability than the others. This paper focuses on detecting cut vertices so that we can either neutralize or protect these critical nodes. Detection of cut vertices is trivial if the global knowledge of the whole system is known but it is very challenging when the global knowledge is missing. In this paper, we propose a completely distributed scheme where every single node can determine whether it is a cut vertex or not. In addition, our design can also confine the detection overhead to a constant instead of being proportional to the size of a network. The correctness of this algorithm is theoretically proved and a number of performance measures are verified through trace driven simulations.