DreamGuard: Efficient Runtime Guardrail for LLM Agents via Risk-Aware World Model

LLM agents that call external tools can trigger irreversible consequences, and existing runtime guardrails mostly react to the apparent safety of the current action without modeling how risk evolves over a trajectory. This creates a blind spot for long-horizon risks where individually benign-looking actions gradually drift the agent toward hazardous states.

DreamGuard addresses this with a risk-aware world model that maintains a compact recurrent latent state over the trajectory and predicts future latent states. From these predictions, it derives immediate-hazard and prefix-risk evidence, fusing multi-horizon signals into intervention decisions before action execution.

Evaluated across four benchmarks and an online guardrail evaluation, DreamGuard outperforms generic, reactive, and proactive guardrail baselines, achieves the best safety-utility trade-off, and maintains an average end-to-end latency of 25 ms per call, making it practical for real-time agent deployments.