RedBench - HarmBench ASR (Direct Request): leaderboard
Safety baseline from RedBench (Dang et al., 2026): the 320 HarmBench test behaviors are sent to the model as plain requests, and Llama-Guard-3-8B judges whether each response is unsafe. The score is the attack success rate in percent; lower is better.
Metric: Attack Success Rate (%; HarmBench test behaviors, Llama-Guard-3-8B judge; raw behavior prompts). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 Nano | 3.44 | #1992 |
| 2 | GPT-4o Mini | 8.75 | #1627 |
| 3 | Gemma 2 9B (IT) | 11.25 | #2275 |
| 4 | Qwen 2.5 7B Instruct | 32.81 | #2583 |
| 5 | Llama 3.1 8B Instruct | 47.5 | #4277 |
| 6 | Ministral-8B-Instruct-2410 | 65 | #4720 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redbench-harmbench-asr-direct-request · How It Works · Data refreshed daily, snapshot 2026-11-12.