HazardAuditor is an execution-grounded framework for training safety guard models for computer-use AI agents, addressing the gap between static prompt/response guard models and existing executable safety platforms that produce evaluation verdicts rather than training signal. Its infrastructure runs heterogeneous agents (Claude Code, Codex, Hermes, and OpenClaw) in controlled environments and normalizes their runtime interactions into a canonical event representation for cross-framework supervision. To fix a training mismatch where longer generated rationales dominate gradient updates, the authors introduce Guard Policy Optimization (GuardPO), which converts deterministic safety outcomes into sequence-level advantages. Across multiple benchmarks and agent systems, HazardAuditor improves guard accuracy by up to 16.5 percentage points over the strongest prior guard model.
