HarnessAudit-Bench: leaderboard
Agent harness safety on 210 benign multi-agent workflow tasks in eight domains with embedded permissions over tools, resources and inter-agent information flow (HarnessAudit-Bench, Liu et al., 2026). The overall harness safety score multiplies each task's safety adherence rate by a weighted mix of task completion (0.7), action validity (0.15) and perturbation stability (0.15), averaged over tasks, 0-1, higher is better. Each entry is a model run inside one harness: OpenClaw, Claude Code or the Codex CLI.
Metric: Overall harness safety score (0-1; task-level safety adherence gate times weighted task completion, action validity and perturbation stability; 210 tasks in eight domains). Source: arxiv.org. Saturation forecast: Around December 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.41 | #91 |
| 2 | GPT-5.4 | 0.32 | #121 |
| 3 | GLM-5V Turbo | 0.31 | #342 |
| 4 | Kimi K2.6 | 0.3 | #173 |
| 5 | Qwen 3.5 Plus | 0.29 | #227 |
| 6 | Claude Sonnet 4.6 | 0.22 | #158 |
| 7 | Claude Opus 4.6 | 0.21 | #100 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harnessaudit-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.