This paper proposes a new method for identifying groups of examinees engaged in collusion that does not require any a priori knowledge of the compromised items. The method, called iterative cluster building (ICB), constructs clusters through an answer similarity framework, sequentially accepting into the cluster examinees whose responses are statistically the most similar to the union of all responses already in the cluster. ICB was evaluated through both simulation and application to a commonly used empirical dataset with known preknowledge. Results from the study were highly encouraging. ICB recovered simulated groups very well, while limiting Type I errors to realistic levels commonly used in operational settings. Application of ICB to the empirical dataset resulted in larger, more homogeneous clusters than have previously been observed