PlugGuard: A Streaming Safeguard for Large Models via Latent Dynamics-Guided Risk Detection

Large models (LMs) are powerful content generators, yet their open‑ended nature can also introduce potential risks, such as generating harmful or biased content. Existing guardrails mostly perform post-hoc detection that may expose unsafe content before it is caught, and the latency constraints further push them toward lightweight models, limiting detection accuracy. In this work, we propose PlugGuard, a novel plug-in framework that enables streaming risk detection within the LM generation pipeline. PlugGuard leverages intermediate LM hidden states through a Streaming Latent Dynamics Head (SLD), which models the temporal evolution of risk across the generated sequence for more accurate real-time risk detection. To achieve reliable streaming moderation in real applications, we introduce an Anchored Temporal Consistency (ATC) loss, ensuring that risk assessments remain consistent with a strict stop-if-harmful policy. Besides, for a rigorous evaluation of streaming guardrails, we also present StreamGuardBench—a model-grounded benchmark featuring on-the-fly responses from each protected model, reflecting real-world streaming scenarios in both text and vision–language tasks. Across diverse models and datasets, PlugGuard consistently outperforms state-of-the-art streaming guardrails (achieving a 22.80% F1 score gain), while using only 20M parameters and adding less than 0.5 ms of per-token latency. The code and StreamGuardBench are released at PlugGuard to facilitate research on streaming guardrails.

LLMs can say harmful things. Current safety checks happen after the text is already written, so bad content might slip through before anyone notices. We built PlugGuard, a lightweight plug-in module that monitors the LLMs' hidden thought process while it is still writing a response, catching risky content word by word as it appears. It learns patterns from how harmful sentences grow, and once it spots danger, it stops the AI immediately. We also created StreamGuardBench, a benchmark that measures safety tools using real live AI outputs instead of pre-written examples. PlugGuard catches more harmful content than existing streaming guardrails, while adding less than half a millisecond of delay per word and using only 20 million trainable parameters.