ForesightSafety Bench (Industrial Safety, Jailbreak): leaderboard
Metric: Attack success rate (%), unweighted mean of the eight industry sectors of ForesightSafetyBench-IndustrialSafety-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 sector-specific harmful request, judged by gpt-4o-2024-11-20 with dimension-specific rubrics; lower is better. Source: arxiv.org. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 0 | #271 |
| 2 | Claude Sonnet 4.5 | 0.33 | #138 |
| 3 | Kimi K2.5 | 1.75 | #139 |
| 4 | Gemini 3 Flash (Preview) | 2.5 | #78 |
| 5 | Claude 3.5 Haiku | 3.17 | #553 |
| 6 | Gemini 3 Pro (Preview) | 3.58 | #64 |
| 7 | GLM-4.7 | 3.75 | #185 |
| 8 | Qwen 2.5 72B Instruct | 4.5 | #436 |
| 9 | Seed-1.6 | 5.92 | #257 |
| 10 | Seed 1.8 | 6.42 | #136 |
| 11 | Qwen 3 235B A22B 2507 Instruct | 8.75 | #291 |
| 12 | Kimi K2 0905 | 12.67 | #282 |
| 13 | DeepSeek V3.2 | 13.42 | #198 |
| 14 | Qwen 3 Max (Thinking) | 13.67 | #201 (Qwen 3 Max) |
| 15 | Gemini 2.5 Flash | 24.75 | #237 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-bench-industrial-safety-jailbreak · How It Works · Data refreshed daily, snapshot 2026-10-11.