HarnessAudit-Bench - Tool Safety Adherence: leaderboard
The tool channel of boundary compliance of HarnessAudit-Bench (Liu et al., 2026): hidden logs audit every tool call of an agent harness on 210 multi-agent workflow tasks in eight domains, and the score is 1 minus the severity-weighted rate of unauthorized, task-irrelevant or role-exceeding tool calls, 0-1, higher is better. Each entry is a model run inside one harness: OpenClaw, Claude Code or the Codex CLI.
Metric: Safety adherence rate, tool channel (0-1; 1 minus the severity-weighted rate of unauthorized, task-irrelevant or role-exceeding tool calls; 210 tasks). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.85 | #91 |
| 2 | Kimi K2.6 | 0.7 | #173 |
| 3 | GLM-5V Turbo | 0.63 | #342 |
| 4 | GPT-5.4 | 0.62 | #121 |
| 5 | Qwen 3.5 Plus | 0.6 | #227 |
| 6 | Claude Sonnet 4.6 | 0.42 | #158 |
| 7 | Claude Opus 4.6 | 0.39 | #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-tool-safety-adherence · How It Works · Data refreshed daily, snapshot 2026-11-12.