The authors consider the detection of a jump of an unknown size at an unknown time of a signal observed in additive white noise. They derive a stochastic differential equation for the conditional density of the signal and use differential-geometric methodology to find finite-dimensional projections that admit recursive implementation. The authors show that the new filter is an adaptive version of Shiryaev-Wonham’s equation for the detection of a priory known changes combined with a modified Kalman filter to estimate the size of the change. Simulation results show excellent performance of the new filter.