Abstract. In this paper we address the problem of inefficiency of exact attack tree computations. We propose several implementation-level optimizations and introduce a genetic algorithm for fast approximate computations. Our experiments show that for attack trees having less than 30 leaves, the confidence level of 89% can be achieved within 2 seconds using this algorithm. The approximation scales very well and attack trees of practical size (up to 100 leaves) can be analyzed within a few minutes.