Hi Archisman,

the leakage is possible, especially if that small protein is a disordered chain. Cut-off values usually refers to globular protein (or molecules), so prolonged spin can result in significant losses of non-globular proteins. To lower the losses you could try to reduce the speed and to periodically mix the suspension in order to homogenize the protein gradient inside the concentrator (less protein next to the membrane means less leakage). However these methods could be time consuming, and cannot completely prevent protein losses. You can also think to other techniques such as lyophilization if your buffer allows this. 

I agree with Emmanuele.  In addition, the cutoff MW advertised for some membranes is a statistical average, rather than an absolute number. For matrix-type filter media, the "pore" sizes are actually a distribution, so some of the openings may be significantly larger than the advertised 3500 MW.