Reproduce “LMSM: LLM Security Framework Inspired by Linux Security Modules” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
XiuYu Zhang, Bonan Ruan, Junfeng Fang, An Zhang, Tat-Seng Chua, Zhenkai Liang
LMSM routes model-internal evidence through operator rules to a separate gate for buffered output release. It evaluates request-keyed policies and gates each response. On Qwen3-4B, it reduced HarmBench attack success rate from 39.20% to 3.32%, while XSTest false-refusal rate rose from 2.40% to 4.40%. It retained 98.14% of matched no-monitoring throughput at 32 active sequences.
Key insights
- Security enforcement for LLMs can be framed as a runtime-mediation problem, drawing inspiration from Linux Security Modules (LSM).
- Separating evidence provisioning (backends), policy evaluation, and output enforcement enables independent evolution and localization of changes within the security stack.
- Maintaining request-keyed policy state is essential to ensure consistent and isolated security decisions in dynamically batched LLM serving environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.