This paper is concerned with the security sliding mode control problem for Markov switching systems suffering from both physical and communication constraints, that is, the actuator saturation and injection attack. With the help of the learning capability of neural networks, the nonlinear upper bound of injection attacks is analyzed and estimated, and the security control strategy based on neural networks is proposed to improve the reliability of control. Considering the transient performance implementation, a time-dependent sliding mode controller is proposed to guarantee the reachability of the specified sliding surface within a given finite time, in which the adaptive rules of neural network parameters are designed to reduce the effects of the considered attack. The arrival and sliding stages are, respectively, analyzed and sufficient conditions are achieved to guarantee the random finite-time boundedness of the closed-loop system. Eventually, a simulation example is given, combined with a genetic algorithm, to illustrate the validity of the proposed adaptive security control method.