ML-Bench (Multilingual Safety) - Attack-Enhanced Accuracy: leaderboard
Detection of refined unsafe queries rewritten by jailbreak optimization (PAIR and AutoDAN), by a guardrail model given only the input text. ML-Bench (Zhao et al., 2026) is a multilingual guardrail benchmark built from 17 regional AI regulations in 14 languages. Share classified unsafe, in percent; higher is better.
Metric: Accuracy (%; share of refined unsafe queries with PAIR/AutoDAN-optimized adversarial suffixes classified unsafe, guardrail given only the input text). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ML-Guard-7B | 90 | |
| 2 | ML-Guard-1.5B | 50 | |
| 3 | Llama-Guard-3-1B | 45 | |
| 4 | DuoGuard-1.5B-transfer | 20 | |
| 5 | Llama-Guard-4-12B | 18 | |
| 6 | Llama-3.1-Nemotron-Safety-Guard-8B-v3 | 18 | |
| 7 | PolyGuard-Qwen | 7 | |
| 8 | Qwen3Guard-Gen-0.6B | 6 | |
| 9 | Llama-Guard-3-8B | 3 | |
| 10 | Qwen3Guard-Gen-8B | 2 |
Interactive version: theaggregate.ai/benchmark?slug=ml-bench-multilingual-safety-attack-enhanced-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.