We consider downlink channel training of a frequency division duplex (FDD) massive multiple-input-multiple-output (MIMO) system when a multi-antenna jammer is present in the network. The jammer intends to degrade mean square error (MSE) of the downlink channel training by designing an attack based on second-order statistics of its channel. The channels are assumed to be spatially correlated. First, a closed-form expression for the channel estimation MSE is derived and then the jammer determines the conditions under which the MSE is maximized. Numerical results demonstrate that the proposed jamming can severely increase the estimation MSE even if the optimal training signals with a large number of pilot symbols are used by the legitimate system.