Reservoir computing enables embedding attractors into random neural networks (RNNs), often generating a “mirror” of a target attractor due to inherent symmetrical constraints in their architecture, which results in multistability. In this work, we report that a simple modulation of a global parameter—the scale of internal connectivity—can induce an attractor-merging crisis accompanied by intermittency, which is characterized by irregular alternation between staying within and escaping from the ruins of embedded attractors. We further reveal its underlying mechanism through a detailed analysis of the phase-space structure, and multiple case studies demonstrate that this bifurcation scenario is intrinsic to a general class of RNNs, independent of training data. This work bridges the gap between machine learning implementations and concepts in nonlinear dynamical systems, contributing to the design of systems that leverage the intrinsic properties of RNNs, as well as to a deeper understanding of biological neural dynamics.