Contextual Reinforcement Learning for Linguistic Intent-Gated Access Control in Production AI Systems

Static policy engines (RBAC/ABAC) cannot adapt to linguistic variability in natural-language access requests. We propose cRL-LIM, a framework integrating contextual reinforcement learning with LLM-based intent parsing: an LLM decomposes requests into structured features processed by a contextual bandit (LinUCB / Thompson Sampling) that fuses semantic intent with runtime context. Across six scenarios totaling$\mathbf{2 7 6, 0 0 0 +}$episodes, LinUCB achieves 99.5 % violation reduction (0.10 % vs. 20.30 % baseline) and maintains this safety guarantee under load stress, adversarial injection, memory poisoning, and policy drift, while static baselines degrade up to 39 %. The learned policy generalizes to unseen oracle policies and held-out request types, confirming that it captures security-relevant features beyond rote memorization.