HazardAuditor runs four agent frameworks in one harness and normalizes their events to train a guard.

arXiv 2609.15134 targets the gap where guard models score static prompts and responses while real computer-use risk comes from runtime behavior across browsers, terminals, filesystems and external services. Its infrastructure runs Claude Code, Codex, Hermes and OpenClaw in controlled environments and flattens their interactions into a canonical event representation. The paper also names a training bug specific to generative guards, where token-level objectives let longer rationales dominate gradients, and fixes it with Guard Policy Optimization, converting deterministic safety outcomes into sequence-level advantages so the verdict is the unit being optimized.