This paper considers the security issue of the reputation systems, which make great use of social networks, to form the reputation scores of a certain participant in the network by collecting ratings from others. To prevent the damage caused by the intentional injection of dishonest ratings, termed as reputation attack, we aim to detect such attack as soon as possible after its occurrence by sequentially observing the rating samples at a single or multiple agents. We propose a discrete probability model to characterize the rating samples based on which the sequential change detection framework can be applied. First for the single-agent case, we develop a sequential attack detector based on the generalized likelihood ratio (GLR) that is robust to a wide range of attacking strategies. Then, for the multi-agent case, where the distributed agents collect ratings and communicate with a central manager to make the decision, we propose a novel multi-agent sequential attack detector that can effectively exploit the different number of attacked agents to increase the detection speed, and exhibits a second-order asymptotic optimality for any given number of attacked agents if the rating distribution under attack can be specified. Finally, we propose a decentralized version of the multi-agent attack detector based on the level-triggered sampling of the local statistic at each agent, which essentially constitutes an adaptive transmission scheme between the distributed agents and the central manager. The decentralized detector incurs only a minor increase in detection delay compared with the centralized counterpart while substantially reduces the communication overhead for attack detection in the multi-agent reputation systems.