Researchers propose Language Model Security Modules (LMSM), a framework that adapts the architecture of Linux Security Modules to protect large language models during deployment. The design separates security concerns into three components: a backend that supplies calibrated evidence from interpretability methods, a versioned policy engine that evaluates rules over request-specific context, and a gating mechanism that controls output authorization. This decouples policy effectiveness from enforcement implementation, letting backends, rules, or scheduling be updated independently without rebuilding the core request-handling logic, and a prototype implementation reduces attack success rates significantly while maintaining throughput comparable to unmonitored serving systems.
