Deep learning provides powerful means to learn from spectrum data and solve complex tasks in 5G and beyond such as beam selection for initial access (IA) in mmWave communications. To establish the IA between the base station (e.g., gNodeB) and user equipment (UE) for directional transmissions, a deep neural network (DNN) can predict the beam that is best slanted to each UE by using the received signal strengths (RSSs) from a subset of possible narrow beams. While improving the latency and reliability of beam selection compared to the conventional IA that sweeps all beams, the DNN itself is susceptible to adversarial attacks. We present an adversarial attack by generating adversarial perturbations to manipulate the over-the-air captured RSSs as the input to the DNN. This attack reduces the IA performance significantly and fools the DNN into choosing the beams with small RSSs compared to jamming attacks with Gaussian or uniform noise.