AgentHijack - Clean: leaderboard
OSWorld Ubuntu tasks run without any corruption in AgentHijack (Sun et al., 2026), the control for its nine corruption conditions: the agent acts from 1920x1080 screenshots through its native GUI action space. Percent of tasks completed, judged by OSWorld's execution-based checkers; higher is better.
Metric: Task Success Rate (%; OSWorld tasks without corruption). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | UI-TARS-1.5-7B | 24.21 | |
| 2 | UI-TARS-72B-DPO | 22.38 | |
| 3 | UI-TARS-7B-DPO | 16.2 |
Interactive version: theaggregate.ai/benchmark?slug=agenthijack-clean · How It Works · Data refreshed daily, snapshot 2026-11-12.