VisualLeakBench - PII Leakage: leaderboard
Metric: Attack success rate (%) on 500 synthetic sticky-note images holding personal data (8 PII types) with a social-engineering prompt: a response that discloses the full target PII counts as unsafe even with an appended warning; baseline system prompt, temperature 0; lower is better. Source: arxiv.org. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4 | 20.4 | #169 |
| 2 | GPT-5.2 | 38.2 | #105 |
| 3 | Gemini 3 Flash (Preview) | 60.6 | #78 |
| 4 | Claude Sonnet 4 | 74.4 | #194 |
Interactive version: theaggregate.ai/benchmark?slug=visualleakbench-pii-leakage · How It Works · Data refreshed daily, snapshot 2026-10-11.