PGuard: a native, streaming, plug-and-play guardrail for multimodal large language models

PGuard is a native, streaming, plug-and-play guardrail for multimodal large language models that uses the backbone model's intermediate representations for token-level safety detection without changing the original inference path. On Qwen3.5-122B-A10B served by SGLang, end-to-end overhead is only 0.82-2.26%. On more than 30 benchmarks covering scenarios such as jailbreak and CBRN risk, it is the best-performing native guardrail, with an average F1 of 0.836 on text queries for Qwen3.5-9B, approaching dedicated external guardrails.