EvoHarmBench (Direct Audit): leaderboard
Metric: Miss rate (%; share of the 5,002 real-world harmful posts in five violation categories that the moderator fails to flag when they are submitted unchanged, with the same category-specific moderation prompts as the iterative protocol). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.5 | 31.7 |
| 2 | Claude Sonnet 4.6 | 36.2 |
| 3 | Gemini 3.1 Pro (Preview) | 37.1 |
| 4 | Qwen 3.6 Plus | 39.5 |
| 5 | Kimi K2.6 | 39.6 |
| 6 | DeepSeek V4 Pro | 41.3 |
| 7 | GLM-5.1 | 43.3 |
Interactive version: theaggregate.ai/benchmark?slug=evoharmbench-direct-audit · How It Works · Data refreshed daily, snapshot 2026-09-26.