Abstract: A prescribed-time fault-tolerant control strategy based on zero-sum differential games is proposed for affine nonlinear systems with actuator faults. The goal is to address the critical issue that current fault-tolerant control methods fail to balance control optimality with explicit time adjustability. As intelligent systems have become more prevalent in such fields as power grids, high-speed railways, and deep-space and deep-sea exploration, safety concerns arising from actuator faults have become increasingly prominent. Traditional methods, such as fuzzy control, adaptive control, and sliding mode control, ensure system stability but struggle to achieve optimal control performance. Moreover, the convergence time typically depends on the initial state or controller parameters, limiting flexibility for user customization. To address this, a differential game framework using an auxiliary function with time and space constraint characteristics was developed. The fault-tolerant control problem is transformed into an adversarial optimization problem, and neural-network technology is employed to mitigate the “curse of dimensionality.” Ultimately, the system achieves both stability and optimal control performance within the prescribed time. For the affine nonlinear system model, an auxiliary function \xi (t) with time–space constraints is introduced. It transforms the original state into a new state constrained by the prescribed time T_\textp and desired accuracy \varepsilon . When the new state function \xi (t) is bounded, the system state converges to the designated accuracy range within a user-specified time. Based on this, the control signal u_\texts and actuator bias fault u_\text0 are modeled as opposing sides of a zero-sum differential game. Using the Nash–Pontryagin maximum–minimum principle, the Hamilton–Jacobi–Isaacs (HJI) equation is derived to solve for the game equilibrium point, yielding the optimal control strategy and boundary values for the bias fault. To address the curse-of-dimensionality problem in solving the HJI equation, an adaptive dynamic programming framework is constructed using a single-critic neural network, which approximates the optimal cost function through online learning, significantly reducing the algorithm complexity. In addition, a real-time adaptive estimation of the fault coefficient matrix is incorporated to enhance the ability of the system to adapt to multiplicative faults. This method not only explicitly defines the convergence time T_\textp but also synchronously optimizes control performance. A simulation using a two-link robotic manipulator demonstrated that, under actuator faults and bias faults, the system can still converge the tracking error within the desired accuracy in the prescribed time. The system exhibits stable tracking performance across different initial states, controller parameters, and prescribed times. Furthermore, the control torque adjusts rapidly during the fault period, thereby suppressing disturbances. Comparative simulations confirmed that the system convergence time is positively correlated with the prescribed time value and independent of the initial state and controller parameters of the system, thereby validating the effectiveness and superiority of the algorithm. The contributions of this study are threefold. First, a zero-sum differential game framework is employed to design an optimal fault-tolerant control strategy, ensuring both the stability of the affine nonlinear system and optimal control performance. This stability is achieved in finite time, in contrast to the asymptotic stability often discussed in differential game theory, which typically assumes infinite time. Second, unlike algorithms that consider only a single fault, in this study, the coexistence of actuator partial faults and bias faults is accounted for because the extreme values of the bias faults are derived based on the game strategy. Furthermore, to address the curse of dimensionality in solving the HJI equation, a single-critic network is constructed using an adaptive dynamic programming algorithm that is further combined with the estimated values of the coefficient fault matrix for integrated updates, thus reducing the complexity of the control algorithm. Third, unlike finite-time and fixed-time control methods, where the convergence time depends on the initial state or controller parameters of the system, the proposed algorithm enables the user to predesign the convergence time, ensuring that the system state converges within a finite time, unaffected by the initial state or controller parameters.