Abstract: The design of secure routing protocol is one of the most critical issues in mobile ad hoc networks(MANET). In MANET, there exists a novel threat where the malicious nodes work in an abnormal manner: they participate in the route discovery phase properly as the regular nodes, while discarding the arriving data packets in the forwarding phase. So far, most of previous research literatures moderately focus on the issues of protecting routing messages, and little work has been put on dealing with this kind of threat. Therefore, it brings great challenges in designing secure routing protocols to react with this threat. In this paper, ASR, a novel adaptive secure routing protocol for mobile ad hoc networks is proposed. In ASR, a node first identifies several paths through the normal routing discovery phase. And then, it selects fault free paths via a reinforcement learning mechanism. Thereafter, a node reputation management mechanism is adopted according to the evidence from reinforcement learning. Finally, in-depth analysis is performed and extensive simulations are conducted in the NS2 simulation environment. Analysis and simulation results show that the proposed adaptive routing protocol is able to work effectively defending itself against data forwarding attacks, and it also performs efficiently in the hostile environment.