TRAP - Privacy: leaderboard

Metric: Privacy score (%): 1 minus the leakage rate under the attack query, sample-weighted over all 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 2035. 22 models tracked.

Top models

#ModelScore
1Gemini 2.5 Pro26.4
2Gemini 2.5 Flash Lite20.4
3GPT-4o Mini13.2
4GPT-512.8
5InternVL3.5-8B12.4
6GPT-5.4 Mini10.4
7Gemini 2.5 Flash10
8GPT-5 Mini10
9Claude Haiku 4.58.8
10Qwen 3 VL 32B8.8
11Claude Sonnet 4.56.6

Interactive version: theaggregate.ai/benchmark?slug=trap-privacy · How It Works · Data refreshed daily, snapshot 2026-09-29.