Dark Patterns Manipulate Web Agents

Published:Dec 28, 2025 11:55
1 min read
ArXiv

Analysis

This paper highlights a critical vulnerability in web agents: their susceptibility to dark patterns. It introduces DECEPTICON, a testing environment, and demonstrates that these manipulative UI designs can significantly steer agent behavior towards unintended outcomes. The findings suggest that larger, more capable models are paradoxically more vulnerable, and existing defenses are often ineffective. This research underscores the need for robust countermeasures to protect agents from malicious designs.

Reference

Dark patterns successfully steer agent trajectories towards malicious outcomes in over 70% of tested generated and real-world tasks.