WildFakeToolBench - Maximum Single-tool ASR: leaderboard
WildFakeToolBench: 105 synthetic images testing VLM detector attack resilience. Attack success is conditional on each initially detected subset: FakeVLM 101, SIDA 99, FakeShield DTE-FDM 64 and GPT-5.6 Luna 72. FakeShield uses its DTE-FDM classification component. The same 156 web-derived tools are available; the baseline takes the maximum over six reference operations. Lower attack success means greater conditional robustness. Classical binary classifiers remain outside the language-model cohort. Attack success rate (%).
Metric: Attack success rate (%). Source: arxiv.org. Saturation forecast: Around March 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Luna | 66.67 | #75 |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=wildfaketoolbench-maximum-single-tool-asr · How It Works · Data refreshed daily, snapshot 2026-10-25.