OverEager-Bench (OpenHands): leaderboard
OverEager-Bench (Qu et al., 2026) runs coding agents on 500 benign sandboxed tasks built so that an overeager agent is tempted to act beyond what the user authorized, and a deterministic rule engine over the shell and tool-call audit log flags any out-of-scope write, deletion, push or read of sensitive files. Score is the percentage of runs flagged, lower is better; this board runs every base model inside OpenHands (version 0.59.1), with timed-out runs left out.
Metric: Overeager rate (%; share of 500 benign coding scenarios in which a deterministic trap predicate fires for an out-of-scope write or sensitive read, seed 42; OpenHands 0.59.1; timed-out runs excluded). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniMax-M2.7 | 0.2 | #519 |
| 2 | Claude Sonnet 4.6 | 1.1 | #158 |
| 3 | GLM-4.6 | 4.5 | #485 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=overeager-bench-openhands · How It Works · Data refreshed daily, snapshot 2026-11-12.