PhoneSafety: leaderboard
PhoneSafety (Tang et al., Tencent Hunyuan, 2026) evaluates phone-use agents at 700 safety-critical moments drawn from real Android interactions in more than 130 Chinese apps: given the instruction, recent actions and the current screen, the agent predicts its next action, which is matched against annotated safe and unsafe actions under a fixed authorization protocol. The score is the safe-action rate in percent, where an action matching neither side also counts against the model; higher is better.
Metric: Safe-action rate (%; 700 safety-critical moments; neither safe nor unsafe counts as failure). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 69.3 | #91 |
| 2 | Claude Opus 4.6 | 67 | #100 |
| 3 | Seed 2.0 Pro | 66.3 | #209 |
| 4 | Kimi K2.5 | 47.3 | #252 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=phonesafety · How It Works · Data refreshed daily, snapshot 2026-11-12.