HazardAuditor runs Claude Code, Codex, Hermes and OpenClaw in one harness and normalizes their events to train a guard model
arXiv 2609.15134, submitted 2026-09-14, targets the gap where existing guard models score static prompts and responses while real computer-use risk emerges from runtime behavior across browsers, terminals, file systems and external services. Its infrastructure runs four heterogeneous agent frameworks in controlled environments and flattens their interactions into a canonical event representation so one guard can learn across all of them. 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.
↳ Follow the thread