Background Attentional bias modification (ABM) aims to reduce anxiety by attenuating bias towards threatening information. The current study incorporated virtual reality (VR) technology and 3-dimensional stimuli with a person-identity-matching (PIM) task to evaluate the effects of a VR-based ABM training on attentional bias and anxiety symptoms. Methods One hundred participants with elevated social anxiety were randomised to four training groups. Attentional bias was assessed at pre- and post-training, and anxiety symptoms were assessed at pre-training, post-training, 1-week follow-up, and 3-month follow-up. Results Change in anxiety did not correlate with change in bias ( r = −0.08). A repeated-measures ANOVA showed no significant difference in bias from pre- to post-ABM, or between groups. For anxiety symptoms, a linear mixed-effects model analysis revealed a significant effect of time. Participants showed reduction in anxiety score at each successive assessment ( p < .001, Nagelkerke's pseudo r 2 = 0.65). However, no other significant main effect or interactions were found. A clinically significant change analysis revealed that 4% of participants were classified as ‘recovered’ at 3-month follow-up. Conclusions A single session of VR-based PIM task did not change attentional bias. The significant reduction in anxiety was not specific to active training, and the majority of participants remained clinically unchanged.