Physical layer security (PLS) relies on avoiding attacks from a physical layer perspective. Beamforming has shown to be an effective alternative for this purpose, allowing the concentration of the resources in the legitimate user. This paper investigates analog beamforming in a massive multiple-input-multiple-output system with an eavesdropper and a jammer. To ensure secure communications, we formulate the secrecy rate maximization problem under the constant modulus constraint of the analog beamforming coefficients. The problem is highly challenging due to the non-convexity and complicated objective function. To overcome the challenges, we propose a deep unfolding architecture that leverages the learning capability of deep neural networks to improve the convergence of the projected gradient ascent (PGA) optimizer. Simulation results show that the proposed deep unfolding beamforming design offers a substantial gain of 33.33% in secrecy rate and a 70% reduction in complexity and run-time concerning the conventional PGA scheme.