Swiss-Bench 003 - Adversarial Security: leaderboard
Adversarial security composite (D8) of Swiss-Bench 003 (2026), a Swiss financial and regulatory test suite in German, French, Italian and English: 0.6 times the personal-data extraction defense score plus 0.4 times the system prompt leakage score, graded by a Qwen3-235B judge with tiered rubrics. Percent; higher is better.
Metric: D8 adversarial security composite (%): 0.6 x PII-Scope score + 0.4 x System Prompt Leakage score, over the Swiss-Bench 003 (SBP-003) adversarial security (D8) items in German, French, Italian and English, zero-shot at provider defaults, judged by Qwen3-235B with tiered rubrics; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 60.7 | #704 |
| 2 | GPT-4o | 55.1 | #751 |
| 3 | Qwen 3.5 Plus | 50.6 | #242 |
| 4 | Claude Sonnet 4 | 44.2 | #410 |
| 5 | MiniMax-M2.5 | 44 | #600 |
| 6 | GLM-5 | 43.2 | #269 |
| 7 | MiMo-V2-Flash | 37.4 | #684 |
| 8 | Gemini 2.5 Flash | 27.9 | #463 |
| 9 | Mistral Large 3 | 23 | #832 |
| 10 | DeepSeek V3.2 | 20.1 | #419 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=swiss-bench-003-adversarial-security · How It Works · Data refreshed daily, snapshot 2026-11-21.