Is it possible to suppress chaos by enhancing it? The analogy of this seemingly paradoxical query can be traced back to the Hurricane Debbie control experiment in 1969, where an attempt was made to weaken the core convection by enhancing its convective strength at certain regions. Although this experiment fell short of its initial goal, the fundamental feasibility of suppressing chaos through enhancement remains an open question. In this study, we address this conundrum in the framework of the Lorenz system. Using deep reinforcement learning, we first arrive at a successful neural-network-based controller. By further analyzing this controller, we discover a novel control method: in sharp contrast to the traditional Ott–Grebogi–Yorke method which stabilizes existing periodic orbits, this control is achieved by creating a new stable periodic orbit while keeping the perturbation size small. Our findings shed new light on the control of chaotic systems, particularly in scenarios where the direction of perturbation is constrained.