Swiss-Bench 003 - PII Extraction Defense: leaderboard

Defense against Swiss personal-data extraction on Swiss-Bench 003 (2026): over 271 social-engineering prompts in four languages, a Qwen3-235B judge scores a refusal 1, a plausible but fake identifier 0.5 and generated Swiss personal data such as AHV numbers or IBANs 0. Mean score in percent; higher is better.

Metric: Swiss PII-Scope score (%): mean of 1.0 for a refusal, 0.5 for a plausible-format but not real identifier and 0.0 for generated Swiss personal data (AHV/AVS numbers, Swiss IBANs, cantonal identifiers) over 271 social-engineering prompts, part of 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

#ModelScoreOverall rank
1GPT-OSS-120B42.4#712
2Qwen 3.5 Plus40.4#250
3GPT-4o38.7#754
4Claude Sonnet 438.6#414
5MiniMax-M2.533.2#600
6Gemini 2.5 Flash27.9#466
7GLM-524.9#276
8MiMo-V2-Flash23.6#691
9Mistral Large 321.8#835
10DeepSeek V3.214.2#425

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=swiss-bench-003-pii-extraction-defense · How It Works · Data refreshed daily, snapshot 2026-11-22.