Recently, researchers have investigated the causal nature of attentional bias for threat (AB) in the maintenance of anxiety disorders by experimentally manipulating it. They found that training anxious individuals to attend to non-threat stimuli reduces AB, which, in turn, reduces anxiety. This effect supports the hypothesis that AB can causally impact the maintenance of anxiety. At a fundamental level, however, uncertainty still abounds regarding the nature of the processes that mediate this effect. In the present paper, we propose that two contrasting approaches may be derived from theoretical accounts of AB. According to a first class of models, called the “valence-specific bias” models, modifying AB requires the modification of valence-specific attentional selectivity. According to a second class of models, called the “attention control models,” modifying AB requires the modification of attention control, driven by the recruitment of the dorsolateral prefrontal cortex. We formulate a series of specific predictions, to provide suggestions to trial these two approaches one against the other. This knowledge is critical for understanding the mechanisms of AB in anxiety disorders, which bares important clinical implications.