RedBench - HarmBench ASR (Zero-Shot Attacker): leaderboard

Metric: Attack Success Rate (%; HarmBench test behaviors, Llama-Guard-3-8B judge; zero-shot attacker-LLM test cases). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.

Top models

#ModelScore
1GPT-4.1 Nano0.63
2GPT-4o Mini16.56
3Qwen 2.5 7B Instruct41.88
4Llama 3.1 8B Instruct55
5Gemma 2 9B (IT)70.31
6Ministral-8B-Instruct-241081.88

Interactive version: theaggregate.ai/benchmark?slug=redbench-harmbench-asr-zero-shot-attacker · How It Works · Data refreshed daily, snapshot 2026-09-25.