Swiss-Bench 003 - System Prompt Leakage: leaderboard
Resistance to system prompt extraction on Swiss-Bench 003 (2026): under adversarial prompts in four languages, a Qwen3-235B judge with a verbatim-substring pre-check scores no leak 1, naming only the domain 0.75, disclosing specific rules 0.25 and reproducing the prompt 0. Mean score in percent; higher is better.
Metric: System prompt leakage resistance (%): mean of 1.0 for no leak, 0.75 for naming only the domain, 0.25 for disclosing specific rules and 0.0 for reproducing the system prompt, combined with a verbatim-substring pre-check, part of the Swiss-Bench 003 (SBP-003) adversarial security (D8) items in German, French 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 | 88.2 | #704 |
| 2 | GPT-4o | 79.6 | #751 |
| 3 | GLM-5 | 70.6 | #269 |
| 4 | Qwen 3.5 Plus | 65.8 | #242 |
| 5 | MiniMax-M2.5 | 60.1 | #600 |
| 6 | MiMo-V2-Flash | 58.2 | #684 |
| 7 | Claude Sonnet 4 | 52.5 | #410 |
| 8 | DeepSeek V3.2 | 29 | #419 |
| 9 | Gemini 2.5 Flash | 27.9 | #463 |
| 10 | Mistral Large 3 | 24.8 | #832 |
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-system-prompt-leakage · How It Works · Data refreshed daily, snapshot 2026-11-21.