arxiv.org/abs/2608.25697

Language Model Security Modules: Turning LLM Interpretability into Runtime Defense

Language Model Security Modules adapts Linux Security Modules concepts to LLM serving, centralizing checks that link interpretability evidence to requests and enforce policies before output. In vLLM tests on Qwen3-4B and Gemma-3-4B under default optimizations, every backend retained at least 96% of native throughput, while one Qwen3-4B probe policy reduced HarmBench attack success from 39.4% to 3.3% and raised XSTest false refusals from 0.8% to 3.1%.


No comments yet

Be the first to share your thoughts.