TRAP - Image Task Accuracy: leaderboard
Task accuracy in percent on the image-document scenarios of TRAP (2026), a benchmark of agents that must use private fields in a tool call while refusing to reveal them when asked. Higher is better.
Metric: Task accuracy (%) on the 200 image-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 July 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 71.5 | #348 |
| 2 | GPT-5.4 Mini | 71 | #389 |
| 3 | Gemini 2.5 Flash | 67.5 | #453 |
| 4 | GPT-5 | 64.5 | #167 |
| 5 | GPT-4o Mini | 64 | #1623 |
| 6 | Gemini 2.5 Pro | 62 | #266 |
| 7 | Claude Sonnet 4.5 | 60 | #250 |
| 8 | Gemini 2.5 Flash Lite | 57 | #958 |
| 9 | Claude Haiku 4.5 | 55.5 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=trap-image-task-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-13.