<p>This study presents a new observer-based security consensus method to address the control problem for nonlinear strict-feedback stochastic multi-agent systems (MASs) under actuator attacks. Different from most existing optimal control methods which rely on the strong assumption of Nash equilibrium existence, here a novel adaptive distributed observer is proposed to remove the assumption. Furthermore, the security issue posed by actuator attacks is tackled by integrating integral sliding-mode (ISM) control into reinforcement learning (RL). The actual backstepping controller comprises an ISM attack compensator that mitigates the effects of these attacks, alongside an RL-based optimal controller that stabilizes the sliding-mode dynamics and achieves optimal performance. Finally, theoretical analysis and practical simulations are performed to validate the effectiveness of the proposed algorithm.</p>