HarnessAudit-Bench - Information-Flow Safety Adherence: leaderboard
The information-flow channel of HarnessAudit-Bench's boundary compliance (Liu et al., 2026): hidden logs audit the messages, forwards and final outputs of an agent harness on multi-agent workflow tasks, and the score is 1 minus the severity-weighted rate of disclosures that the task's communication rules forbid, averaged over the tasks that have such rules, 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, information-flow channel (0-1; mean over tasks with communication rules of 1 minus the severity-weighted leak rate). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.74 | #91 |
| 2 | Kimi K2.6 | 0.63 | #173 |
| 3 | GLM-5V Turbo | 0.62 | #342 |
| 4 | GPT-5.4 | 0.59 | #121 |
| 5 | Qwen 3.5 Plus | 0.57 | #227 |
| 6 | Claude Sonnet 4.6 | 0.53 | #158 |
| 7 | Claude Opus 4.6 | 0.46 | #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-information-flow-safety-adherence · How It Works · Data refreshed daily, snapshot 2026-11-12.