CUAHarm: leaderboard
The model acts as a computer-using agent in an Ubuntu sandbox and is asked, without any jailbreak, to carry out 52 expert-written malicious tasks such as disabling a firewall or installing a backdoor; the number is the share it actually completes, checked by rule-based verifiers on the machine state. Unlike OS-Harm, which uses a judge over traces, success here is verified on the system itself.
Metric: Malicious computer-use tasks completed (%, lower is better). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 42.3 | #167 |
| 2 | Claude 3.5 Sonnet | 51.9 | #746 |
| 3 | Claude Sonnet 4 | 53.8 | #400 |
| 4 | GPT-4o | 57.7 | #741 |
| 5 | Claude 3.7 Sonnet | 59.6 | #458 |
| 6 | Llama 3.3 70B | 65.4 | #1498 |
| 7 | Mistral Large 2 (Jul) | 80.8 | #1155 |
| 8 | Gemini 1.5 Pro | 84.6 | #1049 |
| 9 | Gemini 2.5 Pro | 90.4 | #266 |
Interactive version: theaggregate.ai/benchmark?slug=cuaharm · How It Works · Data refreshed daily, snapshot 2026-11-13.