Generally these programs don't have hard and fast cut-offs. It is mostly subjective on the par tof the user in terms of a trade-off between false-positives and false-negatives. Typically I don't use actual cut-offs but try and combine the data (if using several programs) in a reasonable way to prioritize or rank variants based on the probability of the mutation being damaging. That way you don't miss something obvious to you, but perhaps not obvious to the programs being employed. Some of the programs report the probability of being harmless while others report the probability of being damaging while still others report p-values of significance. The obvious cut-off for p-values is of course 0.05 or 0.1 depending on what you want to allow for a false-discovery rate. For probabilities of being damaging it is far less clear, which is why I advocate using a ranking approach instead of hard filtering with a cut-off.