Individuals with clinical anxiety demonstrate an attention bias toward threatening information, which is thought to be partially driven by heightened amygdala activity to perceived threat. Attention Bias Modification (ABM) is a computer-based treatment that trains attention toward neutral stimuli and away from threatening stimuli. Alterations in initial processing of threat have been linked to ABM responses, but the impact of protracted processing in the aftermath of neutral and threatening information on ABM outcomes has not been well studied. Our study tested whether sustained activity in the amygdala, which occurred after neutral and threatening stimuli had been removed, could predict which individuals would respond well to ABM. Unmedicated anxious individuals underwent a baseline fMRI assessment during performance of a task sensitive to protracted emotional processing. Afterward, they were randomized to complete eight sessions of ABM (n = 38) or a sham training (n = 19). ABM patients who displayed greater sustained bilateral amygdalar response in the aftermath of neutral stimuli displayed the least improvement in self-reported (but not clinician-rated) vigilance symptoms. In contrast, amygdalar response did not predict improvement in sham patients. Results suggest that in certain anxious individuals, the amygdala may have a robust protracted response even to subjectively neutral cues, which could make these individuals a poor fit for ABM because of its focus on repeatedly retraining attention toward neutral cues. Findings may help elucidate neural mechanisms of ABM and promote the identification of a subset of anxious patients who would be good candidates for this intervention.