Researchers Develop DeAction to Detect and Correct Misaligned Actions in AI Agents

Computer-use agents (CUAs) that interact with digital systems autonomously have advanced significantly, but they still struggle with a fundamental problem: actions that drift away from what users actually want them to do.

A new research paper from arXiv tackles this head-on, defining and studying the problem of misaligned actions in CUAs for the first time. Researchers identify two broad sources of misalignment:

• External attacks: Indirect prompt injections and other adversarial techniques that try to hijack an agent's behavior
• Internal failures: Mistakes in the agent's own reasoning that cause it to go off-task

Introducing DeAction and MisActBench#

The team proposes DeAction, a practical guardrail system that works as a safety layer before actions execute. It detects misaligned actions and iteratively corrects them through structured feedback rather than simply blocking them.

To benchmark this work, they created MisActBench—a dataset of realistic agent trajectories with human-annotated labels at the action level, covering three common categories of misalignment found in real-world deployments.

Results#

DeAction outperforms existing baselines substantially:

• Offline evaluation (MisActBench): Over 15% absolute improvement in F1 score compared to baselines
• Online evaluation: Reduces adversarial attack success rate by over 90% while maintaining or even improving task success rates in normal use cases

The approach introduces only moderate latency overhead, making it practical for deployment. This work addresses a critical gap in agent safety—not just blocking bad actions, but understanding and correcting them intelligently.