This paper focuses on designing robust symbol-level precoding (SLP) in the downlink of an overlay cognitive radio (CR) network, where a primary base station (PBS) serving primary users (PUs) and a cognitive base station (CBS) serving cognitive users (CUs) share the same frequency band. When the PBS shares data and perfect channel state information (CSI) with the CBS, an SLP approach which minimizes the CR transmission power and satisfies symbol-wise Safety Margin (SM) constraints of both PUs and CUs, is obtained in a low-complexity quadratic formulation. Then for the case of imperfect CSI from the PBS to CBS, we propose robust SLP schemes. First, with a norm-bounded CSI error model to approximate uncertain channels at the PBS, we adopt the max-min philosophy to conservatively achieve robust SLP constraints. Second, we use the additive quantization noise model (AQNM) to describe the statistics of the quantized PBS CSI, and we employ a stochastic constraint to formulate the problem, where the SM constraints are converted to be deterministic. Simulation results show that the proposed robust SLP schemes help enable PUs to mitigate negative effect of the quantization noise and simultaneously offer CR transmission with significant improvements in energy efficiency compared to non-robust methods.