Reproduce “DECEPTICON: How Dark Patterns Manipulate Web Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Phil Cuvin, Hao Zhu, Diyi Yang
Researchers at Stanford University found that LLM-driven web agents are highly susceptible to deceptive UI designs, known as dark patterns, exhibiting significantly higher vulnerability than human users, with this susceptibility increasing with model size and reasoning capabilities.
Key insights
- LLM-driven web agents are considerably more susceptible to dark patterns (DP rates exceeding 70%) than human users (31-33% DP).
- An "inverse scaling law" was observed, where dark pattern effectiveness *increases* with larger model sizes and greater test-time reasoning compute.
- Common defense mechanisms (in-context prompting, guardrail models) provide only limited and inconsistent protection against various dark pattern types.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.