AgentHijack: Benchmarking Computer Use Agent Robustness to Common Environment Corruptions

Researchers introduced AgentHijack, a benchmark with 9 common environment corruptions to test computer-use agents on desktop tasks.

Even small interface disruptions, such as pop-ups or resolution changes, can sharply degrade agent performance.

They also proposed AgentHijack-Agent, which improves action generation and adds an onlooker to summarize behavior and verify the environment.

The study highlights the need for robustness testing before deploying computer-use agents in real desktop settings.