RouteGuard: A Threat Triage Framework for Specialized Guardrails

The researchers propose RouteGuard, a router-expert framework that triages conversations to specialized expert guardrails for detection by threat type. To evaluate the robustness of guardrail models, the team first built a human-annotated benchmark, GuardZoo, containing 32,460 samples and covering 15 categories of unsafe content. Evaluation on GuardZoo shows that monolithic guardrails suffer from task interference, and that different threat domains require different decision boundaries, making them difficult to compress into a single model. Experiments show that RouteGuard outperforms strong guardrail baselines on fine-grained threat detection, generalizes better under out-of-domain evaluation, and supports modular extension to address emerging threats.