CSSBench - Fraud and Hate: leaderboard
Chinese malicious queries in the fraud and hate domain, disguised with Chinese-specific obfuscation (pinyin, homophones and variant characters, symbol splitting, zero-width characters) in multiple-choice, true/false and open-ended form; CSSBench (Zhou et al., 2026). Attack success rate in percent: the share of responses the Qwen3Guard-Gen-8B judge rates unsafe, greedy decoding; lower is better.
Metric: Attack success rate (%; share of obfuscated malicious fraud and hate queries answered unsafely, Qwen3Guard-Gen-8B judge; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 0.6B | 42 | #8392 |
| 2 | Qwen 3 8B | 44.67 | #1838 |
| 3 | Qwen 3 1.7B | 54 | #5551 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cssbench-fraud-and-hate · How It Works · Data refreshed daily, snapshot 2026-11-12.