Agentic RIS for 6G physical layer security: pre-judgment and tiered policies for autonomous authentication

The forthcoming 6G networks demand scalable, adaptive, and quantum-safe security mechanisms beyond traditional cryptography. In this context, Reconfigurable Intelligent Surface (RIS) offer a transformative potential to enhance Physical Layer Security (PLS) by actively reshaping the wireless environment. While recent Reinforcement Learning (RL)-based methods have advanced RIS optimization towards real-time adaptation, they still face two critical shortcomings: reliance on prior knowledge of legitimate identities, and inflexibility to diverse resource constraints in practical deployments. To overcome these limitations, this paper introduces a novel agentic framework that elevates the RIS to an autonomous security agent. Its core innovations are twofold: first, the agent performs pre-judgment at the RIS side without requiring prior knowledge from upper-layer, actively probing to discriminate between legitimate users and spoofers; second, it employs a tiered policy strategy (myopic, model-based, and model-free) for varying computational and latency budgets. Formulated as a Partially Observable Markov Decision Process (POMDP) integrated with a Gaussian Process Classifier (GPC), our framework enables the RIS agent to learn optimal configurations that amplify the signal of the legitimate user while simultaneously suppressing potential eavesdroppers. Simulation results demonstrate significant reductions in false acceptance rate and authentication latency compared to non-adaptive baselines, validating the framework’s effectiveness for intelligent, resource-aware security.