TRAP - Text Privacy: leaderboard
Metric: Privacy score (%, 1 minus leakage) on the 100 text-document scenarios of 500 document-grounded scenarios (100 text, 200 image, 200 multimodal) over ten private document types; each pairs a task query that needs a tool call with private fields and an attack query eliciting the same fields in natural language; rule-based evaluator, greedy decoding. Source: arxiv.org. Saturation forecast: Around 2031. 22 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash Lite | 13 |
| 2 | Gemini 2.5 Pro | 11 |
| 3 | GPT-5 | 9 |
| 4 | Claude Haiku 4.5 | 7 |
| 5 | InternVL3.5-8B | 4 |
| 6 | GPT-4o Mini | 3 |
| 7 | Gemini 2.5 Flash | 2 |
| 8 | GPT-5 Mini | 2 |
| 9 | Claude Sonnet 4.5 | 1 |
| 10 | GPT-5.4 Mini | 1 |
| 11 | Qwen 3 VL 32B | 1 |
Interactive version: theaggregate.ai/benchmark?slug=trap-text-privacy · How It Works · Data refreshed daily, snapshot 2026-09-29.