The present study was designed to test whether an attentional bias can arise from aversive classical conditioning. Using a differential conditioning paradigm in which slides of angry faces served as conditioned stimuli (CS+/CS-) and electric shock served as unconditioned stimulus (UCS), skin conductance responses (SCRs) of normal subjects (N = 20) were recorded. The effectiveness of the conditioning procedure was proved by differential SCRs to CS+ and CS- slides during the later acquisition trials. During a subsequent extinction phase, sets of three numbers were superimposed on CS+ and CS- slides. subjects were asked to add up these numbers as quickly as possible and to vocalize the results. Vocalization latencies as indexed by chin EMG activity were significantly longer with CS+ than with CS- trials. This result is explained in terms of the attention attracting properties that the CS+ acquires as a result of its pairing with the UCS. It is argued that these properties compete with the attentional resources that are needed for an ongoing task (i.e., addition task). The data suggest that a learning approach to the origins of attentional biases in anxious subjects might be fruitful.