MalOptBench - MOBjailbreak: leaderboard
The same 60 malicious optimization-algorithm requests after the MOBjailbreak method has DeepSeek-V3 rewrite each into a benign-sounding fiction-writing framing. Attack success rate (%) from a GPT-4o judge over three sampled responses per request, lower is safer (MalOptBench, Gu, Wang, Mei, Zhang and Jin, Xidian University, Victoria University of Wellington and Westlake University, 2026).
Metric: Attack Success Rate (%). Source: arxiv.org. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Phi-4 | 88.33 | #1920 |
| 2 | GPT-5 | 96.66 | #163 |
| 3 | GPT-4o (2024-08-06) | 96.66 | #731 |
| 4 | O3 (2025-04-16) | 96.66 | #212 |
| 5 | DeepSeek V3.1 | 98.33 | #500 |
| 6 | Claude Sonnet 4 (20250514) | 98.33 | #386 |
| 7 | Gemini 2.5 Flash | 100 | #452 |
| 8 | DeepSeek V3 | 100 | #705 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 100 | #619 |
| 10 | Command A | 100 | #1292 |
| 11 | Grok 3 Mini | 100 | #535 |
Interactive version: theaggregate.ai/benchmark?slug=maloptbench-mobjailbreak · How It Works · Data refreshed daily, snapshot 2026-11-12.