I think the problem is not the computation of the standard deviation (SD) but the use of the mean and the SD to estimate the probability of an observation several SD away from the mean, under the assumption of normality. If the data have high kurtosis, other distribution, such as the Student's t, should be considered.