Inspired by biological evolution, we consider the following so-called accessibility percolation problem: The vertices of the unoriented $n$-dimensional binary hypercube are assigned independent $U(0, 1)$ weights, referred to as fitnesses. A path is considered accessible if fitnesses are strictly increasing along it. We prove that the probability that the global fitness maximum is accessible from the all zeroes vertex converges to $1-\frac{1}{2}\ln\left(2+\sqrt{5}\right)$ as $n\rightarrow\infty$. Moreover, we prove that if one conditions on the location of the fitness maximum being $v$, then provided $v$ is not too close to the all zeroes vertex in Hamming distance, the probability that $v$ is accessible converges to a function of this distance divided by $n$ as $n\rightarrow\infty$. This resolves a conjecture by Berestycki, Brunet and Shi in almost full generality. As a second result we show that, for any graph, accessibility percolation can equivalently be formulated in terms of first-passage site percolation. This connection is of particular importance for the study of accessibility percolation on trees.