LMSM: LLM Security Framework Inspired by Linux Security Modules
LMSM turns model-internal safety signals into runtime enforcement controls.
The framework separates calibrated evidence, versioned policy rules, and the gate that releases buffered output. Its prototype runs across Hugging Face Transformers and continuously batched vLLM while preserving per-request decisions under scheduler churn. On Qwen3-4B, LMSM-Checkpoint cut HarmBench attack success from 39.20% to 3.32%, with XSTest false refusals rising from 2.40% to 4.40%. Throughput stayed at 98.14% of a matched unmonitored serving path at 32 active sequences. HF Daily Papers' note
The framework separates calibrated evidence, versioned policy rules, and the gate that releases buffered output. Its prototype runs across Hugging Face Transformers and continuously batched vLLM while preserving per-request decisions under scheduler churn. On Qwen3-4B, LMSM-Checkpoint cut HarmBench attack success from 39.20% to 3.32%, with XSTest false refusals rising from 2.40% to 4.40%. Throughput stayed at 98.14% of a matched unmonitored serving path at 32 active sequences. HF Daily Papers' note
score 5