Mobile edge computing (MEC) has emerged as a cutting-edge technique that brings computation and storage resources closer to the edge of the mobile network. However, MEC is vulnerable to be attacked by malicious users. To improve the security of computation tasks and enhance user connectivity, we design a deep reinforcement learning (DRL) network for reconfigurable intelligence surface (RIS)-aided MEC system. Specifically, we jointly optimize the phase shifts at the RIS, tasks offloaded by users and task assignment to maximize the secrecy offloading capacity and minimize energy consumption under different delay requirements of users. Furthermore, a multi-agent twin delayed deep deterministic policy gradient (TD3)-based algorithm is exploited to tackle the non-convex optimization problem. Numerical results validate the feasibility and applicability of our proposed scheme, demonstrating that the proposed scheme significantly improves the security and energy performance of the system compared to the baseline DRL algorithm.