The frozen flux assumption consists in neglecting magnetic diffusion in the core. It has been widely used to compute core flows from geomagnetic observations. Here we investigate the validity of this assumption over the time interval 1980-2005, using high-precision magnetic data from the Magsat, Orsted, and CHAMP satellites. A detectable change of magnetic fluxes through patches delimited by curves of zero radial magnetic field at the core-mantle boundary is associated with a failure of the frozen flux assumption. For each epoch (1980 and 2005), we calculate spatially regularized models of the core field which we use to investigate the change of reversed magnetic flux at the core surface. The largest and most robust change of reversed flux is observed for two patches: one located under St. Helena Island (near 20 degrees S, 15 degrees E); the other, much larger, is located under the South Atlantic Ocean. We next calculate frozen-flux-constrained field models (i.e., pairs of models for epoch 1980 and 2005 having the same flux through patches delimited by curves of zero radial magnetic field), using a penalty method. We find that the frozen flux constraint does not lead to any significant increase of the global misfit. However, applying the constraint leads to a detectable increase of the scalar residuals at satellite altitude in the region of St. Helena, strongly suggesting a local failure of the frozen flux assumption. The observed flux expulsion within the St. Helena patch could result from the formation of a pair of "core spots," as predicted by numerical simulations of the geodynamo.