One of the functions of automatic stimulus evaluation is to direct attention toward events that may have undesirable consequences for the perceiver's well-being. To test whether attentional resources are automatically directed away from an attended task to undesirable stimuli, Ss named the colors in which desirable and undesirable traits (e.g., honest, sadistic) appeared. Across 3 experiments, color-naming latencies were consistently longer for undesirable traits but did not differ within the desirable and undesirable categories. In Experiment 2, Ss also showed more incidental learning for undesirable traits, as predicted by the automatic vigilance (but not a perceptual defense) hypothesis. In Experiment 3, a diagnosticity (or base-rate) explanation of the vigilance effect was ruled out. The implications for deliberate processing in person perception and stereotyping are discussed.