Computer-use agents increasingly operate across browsers, terminals, file systems, and external services, creating safety risks that surface through runtime behavior rather than generated text alone. A new research paper, HazardAuditor, proposes an execution-grounded approach to supervising these agents.
The authors argue that existing guard models, which target static prompts and responses, are poorly suited to agent execution, while existing executable safety platforms produce evaluation verdicts rather than the normalized supervision a guard model needs to learn across heterogeneous agent frameworks.
HazardAuditor addresses both gaps. Its infrastructure runs heterogeneous agents — including Claude Code, Codex, Hermes, and OpenClaw — in controlled environments and normalizes their interactions into a canonical event representation for cross-framework supervision.
The paper also identifies a structural mismatch in token-level post-training objectives for generative guards, where longer rationales dominate gradient updates. Guard Policy Optimization (GuardPO) converts deterministic safety outcomes into sequence-level advantages and normalizes rationale and verdict regions, making the safety decision the effective unit of optimization.
Across multiple benchmarks and heterogeneous computer-use systems, HazardAuditor improves accuracy by up to 16.5 percentage points over the strongest prior guard. Code, models, and evaluation artifacts are slated to be released on the project page.