The attentional blink (AB) is a transient attentional deficit that occurs when two stimuli that must both be detected are presented within an interval of less than 500 ms. Event-related potential (ERP) investigations have suggested that the AB affects a specific component, the P3, which is suppressed when targets are blinked. In view of the link between the P3 and working memory, it has been suggested that the AB might be due to the inability of the blinked target to access working memory. Interestingly, it seems that faces, due to their saliency, might escape the AB effect when cross-category detection is required (i.e., when the targets are composed of faces versus other categories of stimuli). In the present study we investigated this phenomenon in an event-related potential (ERP) study using upright and inverted faces as targets. In a first task, the participants were asked to identify two successive targets, the first composed of geometric shapes and the second of upright or inverted faces. A second control task, identical to the first was also performed, in which only the second targets had to be identified in order to compare ERPs. ERPs and scalp topographies of physically identical sequences of events, differing only by the attentional involvement, were thus compared. Behavioural results showed that faces indeed escape the AB while inverted faces do not. However, the electrophysiological findings showed that when attention was engaged in a previous stimulus (at the shortest lag times), both upright and inverted faces showed a decreased amplitude in the 150-260 ms time period, in addition to a lower P3. At longer lags, when the AB was no longer observed, no ERP differences were found. Our data demonstrate that, although faces escape the attentional blink, previous attentional involvement occurs much earlier than described for other categories of stimuli. This suggests that faces are subjected to an early selection which might allow rapid re-allocation of attention to the stimulus if it is deemed meaningful.