Background: Models of anxiety disorders and the rationale of exposure therapy (ET) are grounded on classical fear conditioning. Yet, it is unclear whether lower fear ratings of conditioned safety versus threat cues and corresponding neural markers of safety-learning and/or fear inhibition assessed before treatment would predict better outcomes of behavioral exposure. Methods: Sixty-six patients with spider phobia completed pre-treatment clinical and experimental fear conditioning assessments, one session of virtual reality ET, a post-treatment clinical assessment, and a 6-month follow-up assessment. Tilted Gabor gratings served as conditioned stimuli (CS) that were either paired (CS+) or remained unpaired (CS-) with an aversive phobia-related and phobia-unrelated unconditioned stimulus (UCS). CS+/CS- differences in fear ratings and magnetoencephalographic event-related fields (ERFs) were related to percentual symptom reductions from pre- to post-treatment, as assessed via spider phobia questionnaire (SPQ), behavioral avoidance test (BAT), and remission status at 6-month follow-up. Results: We observed no associations between pre-treatment CS+/CS- differences in fear ratings and any treatment outcome. CS+/CS- differences in source estimations of ERFs revealed that higher CS- activity in bilateral dorsolateral prefrontal cortex (dlPFC) was related with SPQ- and BAT-reductions. Associations between CS+/CS- differences and treatment outcomes were also observed in left ventromedial prefrontal cortex (vmPFC) regions, which additionally revealed associations with the follow-up remission status. Conclusions: Results provide initial evidence that neural pre-treatment CS+/CS- differences may hold predictive information regarding outcomes of behavioral exposure. Our findings highlight a key role of neural responses to safety cues with potentially inhibitory effects on affect-generating structures during fear conditioning.