ForesightSafety Bench (Environmental AI Safety, Jailbreak): leaderboard
Metric: Attack success rate (%), unweighted mean of the seven environmental-harm dimensions of ForesightSafetyBench-EnvAI-O: each harmful request is wrapped in the paper's templated role-play jailbreak prompts from JailbreakChat (AIM, DAN and developer-mode templates), and the rate is taken over all attacked prompts; a success is a response that complies with the environmentally harmful request, judged by gpt-4o-2024-11-20 with dimension-specific rubrics; lower is better. Source: arxiv.org. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 0 | #271 |
| 2 | Grok 4 | 0 | #169 |
| 3 | GPT-4o (2024-11-20) | 0 | #369 |
| 4 | Qwen 3 Max (Thinking) | 0 | #201 (Qwen 3 Max) |
| 5 | Claude Sonnet 4.5 | 1.43 | #138 |
| 6 | Grok 4 Fast (Non-reasoning) | 2.5 | #242 (Grok 4 Fast) |
| 7 | GPT-5.2 | 3.93 | #105 |
| 8 | Kimi K2.5 | 6.25 | #139 |
| 9 | GLM-4.7 | 8.39 | #185 |
| 10 | Qwen 2.5 72B Instruct | 8.57 | #436 |
| 11 | Claude 3.5 Haiku | 9.82 | #553 |
| 12 | Gemini 3 Pro (Preview) | 10 | #64 |
| 13 | DeepSeek V3.2 | 10.36 | #198 |
| 14 | Gemini 3 Flash (Preview) | 11.07 | #78 |
| 15 | Kimi K2 0905 | 14.64 | #282 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-bench-environmental-ai-safety-jailbreak · How It Works · Data refreshed daily, snapshot 2026-10-11.