When humans grasp and manipulate objects there are many choices to make, such as where to place the digits or how much force each digit should apply. This problem is highly unconstrained as infinitely many different combinations of finger positions and force distributions lead to stable grasps. This is due to the many redundancies at different levels of the sensorimotor system. In this paper, we investigate the strategy used by humans in distributing finger positions and forces while a hand-held object was perturbed by force and torque in a predictable or unpredictable fashion. Our results revealed that there was a substantial systematic variability among participants' initial placement of the digits on the object. However, within participants' digit placement was rather stereotypical. Moreover, the normal forces applied by the digits co-varied with their initial horizontal and vertical placements. Importantly, we recorded an effect of the horizontal and vertical shift between the thumb and the virtual finger positions on the grip force. Principal component analysis revealed that more than 95% of the digit force variance was accounted by the first two components. Finally, participants learned to compensate the external torque within the first perturbations within each trial during the holding phase. We propose that digit forces were modulated online based on self-chosen digit locations during the holding phase in order to successfully compensate sudden external torques.