The Empirical Rule states that, when the distribution is bell-shaped, around 95% of all observations will be within 2 standard deviations from the mean, which means from μ - 2σ to μ + 2σ.
The binomial distribution is a bell-shaped distribution with:
μ = np and
σ = √(np(1-p))
Therefore, 95% of the outcomes will be in the interval from np-2√(np(1-p)) and np+2√(np(1-p))
Any observation that lies outside this interval (which means greater than np+2√(np(1-p)) or less than np-2√(np(1-p)) ) occurs less then 5% of the times and therefore can be considered "unusual".