The current practice in dam safety studies is to use standardized breach parameters to produce a breach outflow hydrograph and to map the flood plain. Multiple standards define the width, height, side slope, and formation time of the breach based on a sample of historical dam failures. However, the use of standardized parameters introduces uncertainties in the breach outflow hydrograph which are not estimated in current dam safety studies and may lead to a poor assessment of the dam failure consequences. New probabilistic methods based on a database of historical failures are proposed in the literature to overcome this problem. In this paper, the case study of the 1996 Cut-Away dyke failure on the Ha! Ha! river in the Saguenay region in Quebec is used to demonstrate how the use of standards or probabilistic methods based on historical failures can lead to a poor estimate of the breach outflow hydrograph. A new methodology is proposed to retrieve the breach parameters of historical failures using a gradient-based optimization algorithm and reservoir water levels. This methodology is applied to the 1996 Cut-Away dyke failure. The results show that for this case study, standards and probabilistic methods based on historical events cannot properly predict breach parameters, leading to an underestimation of 85% and 45% of the peak discharge.