DreamGuard: Proactive LLM Agent Guardrail with Risk-Aware World Model.

As Large Language Model (LLM) agents increasingly interact with external tools and real-world systems, the potential for unsafe actions leading to irreversible consequences grows. Existing runtime guardrails primarily react to immediate risks, assessing the safety of a single proposed action without explicitly modeling how risk evolves over an entire trajectory. This reactive approach creates a blind spot for long-horizon risks, where a series of seemingly benign actions can cumulatively lead to hazardous states. To address this, researchers propose DreamGuard, a proactive guardrail system built around a novel risk-aware world model. This world model maintains a compact recurrent latent state that tracks the agent's trajectory and predicts future latent states. From these predictions, DreamGuard can derive both immediate hazard evidence and prefix-risk evidence, offering a multi-horizon view of potential dangers. By fusing these diverse risk signals, DreamGuard makes informed intervention decisions before an action is executed. Experiments across four benchmarks and an online evaluation demonstrate that DreamGuard surpasses generic, reactive, and other proactive guardrail baselines, achieving a superior safety-utility trade-off. Crucially, it maintains an average end-to-end latency of just 25 ms per call, making it efficient for real-time deployment.