AgentHijack: leaderboard
Computer-use robustness benchmark (AgentHijack; Sun et al., 2026) built on the OSWorld Ubuntu task list: the agent acts from 1920x1080 screenshots through its native GUI action space, once without corruption and once under each of nine configurable environment corruptions, and OSWorld's execution-based checkers decide success. The score is the mean task success rate over the clean run and the nine corrupted runs, in percent; higher is better.
Metric: Task Success Rate (%; mean of the clean run and the nine corruptions). 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 | 18.74 | |
| 2 | UI-TARS-72B-DPO | 16.96 | |
| 3 | UI-TARS-7B-DPO | 13.14 |
Interactive version: theaggregate.ai/benchmark?slug=agenthijack · How It Works · Data refreshed daily, snapshot 2026-11-12.