Biased attention toward threatening facial expressions is an important maintaining and possibly aetiological factor for social anxiety. However, little is known about the underlying mechanisms. To develop our understanding of this threat bias, the relative contributions of top-down attention, bottom-up attention, and selection history were differentiated across four studies. In Study One, the roles of top-down attention, bottom-up attention, and selection history were tested in an unselected sample using a modification of the dot-probe task, in which participants were cued to attend to a happy or angry face on each trial. Results showed that attentional orienting toward facial expressions was not exclusively driven by bottom-up attentional capture as some previous theories suggest; but instead, participants could shift attention toward emotional faces in a top-down manner. This effect was eliminated when the faces were inverted, demonstrating that top-down attention relies on holistic face processing. Study One found no evidence of selection history (i.e., no improvement on repeated trials or blocks of trials in which the task was to orient to the same expression). Study Two tested whether this ability to use top-down attention to orient to emotional faces is impaired for individuals with social anxiety. Using the same task as Study One, Study Two found that participants with higher levels of social anxiety were selectively impaired in attentional shifting toward a cued happy face when it was paired with an angry face, but not when paired with a neutral face. These results indicate that high social anxiety is associated with deficits in top-down control of attention, which are selectively revealed in the presence of non-task-relevant threat. The results of Study Two could be explained by bottom-up attention to threat or a top-down set for threat that could not be overcome by the instruction to attend to a happy face. To test this, Study Three utilised a modified dot-probe task in which participants were presented with an upright face paired with an inverted face (displaying a disgust or neutral expression) and engagement with and disengagement of attention from threatening faces were measured separately. The task was performed under no, low, and high working-memory load conditions. Since working-memory load draws on the same resources as top-down attention, interference from increasing working-memory load on attentional orienting would point to a role for top-down attention. Social anxiety was not associated with delayed disengagement from threat. However, surprisingly, high social anxiety was associated with an engagement bias away from threat, while low social anxiety was associated with a bias toward threat. These results were unaffected by the working-memory load manipulation. However, some methodological issues were identified with the study. Study Four overcame these methodological issues by using a paired angry and neutral face under no, low and high working-memory load conditions. Higher levels of social anxiety were associated with increased engagement with threat under no-load, but not under low- and high-load conditions. Thus, this body of research provides evidence that social anxiety is associated with an engagement bias to threat, which is driven by top-down attention.