MobileWorldSafety - Attack Success Rate: leaderboard
MobileWorldSafety (Shanghai AI Laboratory, 2026) measures the safety of LLM-powered GUI agents on real Android apps under environmental injection attacks, over 142 risk tasks. The attack success rate is the share of scorable tasks in which the agent executes or partially executes the injected risk action, in percent, lower is better.
Metric: Attack success rate (%; share of scorable tasks in which the agent executes or partially executes the injected risk action; environmental injection attacks over 142 real Android-app tasks; no extra safety system prompt). Source: arxiv.org. Saturation forecast: Around November 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 47.5 | #244 |
| 2 | Kimi K2.5 | 49.2 | #252 |
| 3 | Qwen 3.5 397B A17B | 51.6 | #242 |
| 4 | Gemini 3 Pro | 66.9 | #135 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mobileworldsafety-attack-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-12.