When Actions Go Off-Task: Detecting and Correcting Misaligned Actions in Computer-Use Agents

TL;DR

This paper introduces MisActBench and DeAction, a framework for detecting and correcting misaligned actions in computer-use agents, improving safety and reliability by addressing both external and internal causes of misalignment.

Contribution

It is the first comprehensive study on misaligned action detection in CUAs, proposing a benchmark and a universal guardrail for real-time correction.

Findings

DeAction outperforms baselines by over 15% in F1 score on MisActBench.

Reduces attack success rate by over 90% in adversarial settings.

Maintains or improves task success rate in benign environments.

Abstract

Computer-use agents (CUAs) have made tremendous progress in the past year, yet they still frequently produce misaligned actions that deviate from the user's original intent. Such misaligned actions may arise from external attacks (e.g., indirect prompt injection) or from internal limitations (e.g., erroneous reasoning). They not only expose CUAs to safety risks, but also degrade task efficiency and reliability. This work makes the first effort to define and study misaligned action detection in CUAs, with comprehensive coverage of both externally induced and internally arising misaligned actions. We further identify three common categories in real-world CUA deployment and construct MisActBench, a benchmark of realistic trajectories with human-annotated, action-level alignment labels. Moreover, we propose DeAction, a practical and universal guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. DeAction outperforms all existing baselines across offline and online evaluations with moderate latency overhead: (1) On MisActBench, it outperforms baselines by over 15% absolute in F1 score; (2) In online evaluation, it reduces attack success rate by over 90% under adversarial settings while preserving or even improving task success rate in benign environments.