We study the detection of misbehavior in a Gaussian relay system, where the source transmits information to the destination with the assistance of an amplify-and-forward relay node subject to unreliable channel state information (CSI). The relay node may be potentially malicious and corrupt the network by forwarding garbled information. In this situation, misleading feedback may take place, since reliable CSI is unavailable at the source and/or the destination. By classifying the action of the relay as detectable or undetectable, we propose a novel approach that is capable of coping with any malicious attack detected and continuing to work effectively in the presence of unreliable CSI. We demonstrate that the detectable class of attacks can be successfully detected with a high probability. Meanwhile, the undetectable class of attacks does not affect the performance improvements that are achievable by cooperative diversity, even though such an attack may fool the proposed detection approach. We also extend the method to deal with the case in which there is no direct link between the source and the destination. The effectiveness of the proposed approach has been validated by numerical results.