ML-Bench (Multilingual Safety, Policy-Conditioned) - Attack-Enhanced Accuracy: leaderboard
Detection of refined unsafe queries rewritten by jailbreak optimization (PAIR and AutoDAN), by a guardrail model also given the risk category and safety rule. 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 also given the risk category and safety rule). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ML-Guard-7B | 92 | |
| 2 | gpt-oss-safeguard-20b | 44 |
Interactive version: theaggregate.ai/benchmark?slug=ml-bench-multilingual-safety-policy-conditioned-attack-enhanced-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.