If you’re reducing the expected risk of existential disaster by a larger amount, you’re in expectation net saving lives rather than net killing. If all options involve existential risk, including doing nothing, then all one can do is pick the option with the lowest.
Any other possible effects don’t negate that you’re killing six million people when you’re going ahead with a potentially UnFriendly AI.
If you’re reducing the expected risk of existential disaster by a larger amount, you’re in expectation net saving lives rather than net killing. If all options involve existential risk, including doing nothing, then all one can do is pick the option with the lowest.