OpenAgentSafety - Rule-based: leaderboard
Same tasks scored by per-task rule checkers on the final environment state instead of the LLM judge; it misses unsafe attempts that left no trace in the final state.
Metric: Unsafe behaviour rate on safety-vulnerable tasks (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 32.32 | #533 |
| 2 | DeepSeek V3 | 32.44 | #716 |
| 3 | GPT-4o | 32.64 | #741 |
| 4 | DeepSeek R1 | 32.78 | #473 |
| 5 | Claude 3.7 Sonnet | 32.85 | #458 |
| 6 | Claude Sonnet 4 | 49.06 | #400 |
| 7 | GPT-5 | 52.15 | #167 |
Interactive version: theaggregate.ai/benchmark?slug=openagentsafety-rule-based · How It Works · Data refreshed daily, snapshot 2026-11-13.