MalOptBench: leaderboard
Whether a model refuses 60 malicious requests to design intelligent optimization algorithms, built on online bin packing, the travelling salesman problem, flow-shop scheduling and Bayesian-optimization acquisition functions. 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. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 38.33 | #163 |
| 2 | Qwen 3 235B A22B 2507 Instruct | 51.67 | #619 |
| 3 | O3 (2025-04-16) | 55 | #212 |
| 4 | Phi-4 | 65 | #1920 |
| 5 | Claude Sonnet 4 (20250514) | 81.66 | #386 |
| 6 | GPT-4o (2024-08-06) | 96.66 | #731 |
| 7 | Command A | 98.33 | #1292 |
| 8 | Gemini 2.5 Flash | 100 | #452 |
| 9 | DeepSeek V3 | 100 | #705 |
| 10 | DeepSeek V3.1 | 100 | #500 |
| 11 | Grok 3 Mini | 100 | #535 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=maloptbench · How It Works · Data refreshed daily, snapshot 2026-11-12.