Membership Inference Attacks (MIAs) are widely used to audit the privacy of Artificial Intelligence (AI)-powered systems, which, despite their indisputable success across a myriad of tasks, are known to expose their training data. Moreover, recent work has shown that MIAs should account for disparities arising from inherent random processes in their functioning when evaluating data exposure. In this paper, we analyze the reliability and completeness of attacks on unsupervised MIAs: when the attacker has a target set that overlaps with some of the training data but does not know which instances are in the target set, often resorting to clustering methods to distinguish between members and non-members. Through a series of experiments, we show how to adapt the framework of Coverage, and Stability to the unsupervised setting, assessing the influence of the evaluated set and the clustering seed on the privacy estimation being measured. We explore unsupervised attacks tailored for the DocVQA task and present evidence that applying the same attack to subsamples of the target set can yield different predictions, which can be harnessed to increase true positive rates.