Search papers, labs, and topics across Lattice.
This paper introduces DreamGuard, a proactive runtime guardrail for large language model (LLM) agents that utilizes a risk-aware world model to predict future latent states and assess long-horizon risks. By maintaining a compact recurrent latent state, DreamGuard effectively fuses immediate-hazard and prefix-risk evidence to inform intervention decisions prior to action execution. Experimental results demonstrate that DreamGuard significantly outperforms existing reactive and proactive guardrail baselines, achieving superior safety-utility trade-offs while maintaining low latency.
DreamGuard achieves a groundbreaking safety-utility balance by predicting long-term risks, outperforming traditional guardrails that only react to immediate threats.
As large language model (LLM) agents increasingly invoke external tools and interact with real-world systems, unsafe actions may cause irreversible consequences on external states, user data, and downstream services. Recent runtime guardrails mitigate such risks by checking proposed actions before execution, but many remain reactive: they primarily assess the apparent safety of the current action, lacking an explicit model of how risk evolves across the trajectory. This limitation creates a critical blind spot for long-horizon risks, where individually benign-looking actions can gradually drift the agent toward hazardous states. In response, we propose DreamGuard, a proactive guardrail for LLM agents built around a risk-aware world model. The world model maintains a compact recurrent latent state over the trajectory and predicts future latent states from which DreamGuard derives immediate-hazard and prefix-risk evidence. It then fuses these multi-horizon signals into intervention decisions before execution. Experiments across four benchmarks and an online guardrail evaluation show that DreamGuard outperforms generic, reactive, and proactive guardrail baselines, achieves the best safety-utility trade-off among evaluated guardrails, and maintains an average end-to-end latency of 25 ms per call.
