The vigilance-avoidance hypothesis of selective attention assumes that socially anxious persons initially direct their attention toward fear-related stimuli and subsequently avoid these social stimuli to reduce emotional distress. New technical developments provide tools to implicit measure overt attention on fear-related stimuli via eye-tracking in ecological valid virtual environments presented via a head-mounted display. We examined in 27 low (LSA) and 26 high socially anxious (HSA) individuals fear ratings, physical behavior (duration of approach), hypervigilance (time to first fixation), and attentional avoidance (count of fixations) toward virtual female and male agents (CS) during social fear conditioning (SFC) and extinction in virtual reality (VR). As hypothesized, generally SFC was successfully induced and extinguished concerning the fear ratings. Our findings partly support the vigilance-avoidance hypothesis as HSA directed especially at the first half of the fear acquisition their initial attention more at CS+ than CS- agents, and avoided subsequently the CS+ more than the CS- agents during the fear acquisition. In contrast, in LSA participants initial and sustained attention did not differ between CS+ and CS- agents during fear acquisition. We conclude that HSA individuals guide their initial attention to emotionally threatening stimuli and subsequently avoid the threatening stimuli to possibly reduce their emotional distress, whereas LSA individuals regulate themselves less in their (fear) responses during SFC. Measuring implicit gaze behavior within a well-controlled virtual environment is an interesting innovative tool to in deeply investigate the impact of attention on emotional learning processes.