HazardAuditor Claims 16.5-Point Gain Over Prior Agent Guards
TL;DR
- HazardAuditor reports up to 16.5 percentage points accuracy improvement over the strongest prior guard on computer-use agent safety benchmarks.
- The framework runs Claude Code, Codex, Hermes and OpenClaw in controlled environments and normalizes their interactions into a canonical event representation.
- Its Guard Policy Optimization objective makes the safety decision, not the individual token, the effective unit of learning.
HazardAuditor, a new agent-safety framework, claims up to 16.5 percentage points of accuracy improvement over the strongest prior guard on computer-use agents. The arXiv preprint runs heterogeneous agents in controlled environments and normalizes their behavior into what the authors call a canonical event representation for cross-framework supervision.
The systems it covers: Claude Code, Codex, Hermes, and OpenClaw. The paper's argument is that static prompt-response guards miss what actually happens at runtime, since agents introduce 'safety risks that emerge through runtime behavior rather than generated content alone.'
The training method is Guard Policy Optimization, or GuardPO. In the authors' words, it works by 'converting deterministic safety outcomes into sequence-level advantages and normalizing rationale and verdict regions, making the safety decision the effective unit of optimization.' That reframing is the engine behind the 16.5-point number.
The abstract publishes no per-benchmark or per-agent accuracy breakdowns. Code, models, and evaluation artifacts are promised at yunhao-feng.github.io/HazardAuditor/.
Originally reported by arxiv.org
Read the original article →Original headline: HazardAuditor Paper Trains Agent Guards on Executable Threats, Beats Baselines by 16.5 Points