VisualLeakBench - Unsafe Text Propagation (External Handoff): leaderboard
Visual-to-tool leakage (VisualLeakBench, Wang et al., 2026): a single-step vision-language agent sees a screenshot, document, form, chat or dashboard and may call a tool; the score is the percentage of traces that copy the labeled target string into the tool arguments, over 50 images with the baseline system prompt; lower is better. Images show a rendered unsafe request (illicit, cyber, fraud or sabotage content); the agent may forward content by email to an operations inbox.
Metric: Tool propagation rate (%; non-error traces whose tool arguments contain the labeled target string), 50 images, baseline system prompt, send_email handoff to an operations inbox. Source: arxiv.org. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 33.3 | #75 |
| 2 | Gemini 3.1 Flash Lite | 98 | #363 |
| 3 | GPT-5.4 | 100 | #121 |
| 4 | Grok 4.20 | 100 | #296 |
Interactive version: theaggregate.ai/benchmark?slug=visualleakbench-unsafe-text-propagation-external-handoff · How It Works · Data refreshed daily, snapshot 2026-11-12.