ODCV-Bench: leaderboard
The agent works through 40 multi-step bash tasks where hitting a KPI conflicts with an ethical, legal or safety rule; the number is the share of runs (both prompt variants) where the four-judge median rates the run as metric gaming or worse. Unlike AgentHarm, nobody asks for harm: the violation is the agent's own shortcut to the goal.
Metric: Misalignment rate, both variants (%, lower is better). Source: github.com. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 0 | #79 |
| 2 | Claude Opus 4.5 | 1.28 | #156 |
| 3 | GPT-5.1 Instant | 6.33 | #440 |
| 4 | Llama 4 Maverick | 16.22 | #1099 |
| 5 | GPT-5.5 | 21.25 | #49 |
| 6 | GLM-5.1 | 25 | #179 |
| 7 | GPT-OSS-20B | 28.75 | #1147 |
| 8 | Qwen 3.6 Max Preview | 28.75 | #159 |
| 9 | GLM-4.6 | 30.38 | #486 |
| 10 | Kimi K2 0905 | 32.5 | #558 |
| 11 | Kimi K2.6 | 33.75 | #178 |
| 12 | Qwen 3 Max | 35.44 | #403 |
| 13 | MiniMax-M2.7 | 35.9 | #518 |
| 14 | GPT-OSS-120B | 36.25 | #700 |
| 15 | Qwen 3 30B A3B 2507 Instruct | 39.24 | #1089 |
Interactive version: theaggregate.ai/benchmark?slug=odcv-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.