With the advancement of Artificial Intelligence (AI) and Explainable AI (XAI), Deep Learning (DL) algorithms become increasingly vulnerable to adversarial machine learning (AML) threats. One of the widely used DL-based applications in wireless communication is the Automatic Modulation Classification (AMC). In this paper, we investigate XAI methods for designing AML attacks in DL-based AMC systems. Specifically, we discuss four AML tactics, namely, untargeted causative, targeted evasion, untargeted evasion, and targeted backdoor attacks. We present new specific explanation methods used to launch these attacks, which are customized versions of Saliency Map, Integrated Gradient, the SHapley Additive exPlanations (SHAP), and the enhanced Gradient-weighted Class Activation Mapping (Grad-CAM++) techniques. These explanations identify the most contributory data points to the DL-based AMC classifier. It allows an attacker to perturb these identified data samples using the Fast Gradient Sign Method (FGSM). We compare these new XAI-guided attacks with traditional state-of-the-art attacks on DL-based AMC. Our research shows that the XAI-guided threats outperform (negatively) the traditional ones in terms of Robust Accuracy (RA) metric, Attack Success Rate (ASR), and Fooling Rate (FR), but they exhibit higher computational complexity. However, XAI-guided AML attacks require fewer perturbed signals, which justifies the exhibited computational cost.