OpenAgentSafety: leaderboard
An OpenHands agent with a real shell, file system, browser and simulated colleagues handles 356 multi-turn tasks with benign, ambiguous or adversarial intent; the number is the share of tasks where a GPT-4.1 judge finds unsafe behaviour, counted over tasks where the agent reached a risky state. Unlike AgentHarm the tools are real and many requests look harmless.
Metric: Unsafe behaviour rate on safety-vulnerable tasks (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 49.06 | #400 |
| 2 | Claude 3.7 Sonnet | 51.2 | #458 |
| 3 | GPT-5 | 52.58 | #167 |
| 4 | DeepSeek V3 | 62.23 | #716 |
| 5 | DeepSeek R1 | 62.78 | #473 |
| 6 | GPT-4o | 65.8 | #741 |
| 7 | O3 Mini | 72.73 | #533 |
Interactive version: theaggregate.ai/benchmark?slug=openagentsafety · How It Works · Data refreshed daily, snapshot 2026-11-13.