Research map / Multi-agent systems and games / LLM agents
How Dark Patterns Manipulate Web Agents
Phil Cuvin, Hao Zhu, Diyi Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that dark patterns on websites are highly effective at steering web agent trajectories away from user goals, and introduces DECEPTICON, an environment for testing individual dark patterns in isolation.
Problem
Web agents may be manipulated by deceptive UI designs (dark patterns) that push them to take actions misaligned with their goals, creating a robustness risk.
Approach
An environment (DECEPTICON) that isolates individual dark patterns so their effect on agent trajectories can be measured separately.
Key results
- Dark patterns are reported as highly effective in steering web agent trajectories away from user goals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| insert dark patterns into web tasks | web agents (GPT-4o) · Decepticon generated split | task success rate: 99.4 -> 19.6 | Hurt |
| compare agents with humans on dark pattern effectiveness | LLM web agents · Decepticon generated split | DP (%): 31.0 (human) vs 59.6-78.5 (agents) | Hurt |
| scale model size from 3B to 72B | Qwen2.5-VL · Decepticon generated split | DP (%): 38.5 -> 73.7 | Hurt |
| increase thinking tokens from 256 to 16,384 | Gemini-2.5-Flash · Decepticon generated split | DP (%): 37.6 -> 71.2 | Hurt |
| raise reasoning effort low to high | OpenAI o3 · Decepticon generated split | DP (%): 67.2 -> 70.2 | No clear effect |
| use coordinate-based agent instead of set-of-marks scaffold | web agents · Decepticon generated split | DP (%): 74.7 -> 68.3 | No clear effect |
| add in-context prompting about dark patterns (ablation) | Simple agents · Decepticon generated split | DP (%): 78.5 -> 59.6 (GPT-4o) | Improved |
| add guardrail model that flags dark patterns (ablation) | Simple agents · Decepticon generated split | DP (%): average reduction 28.6% | Improved |
| add guardrail model that flags dark patterns (ablation) | Simple agents · Decepticon generated split | task success rate: 23.4 -> 42.6 (ICP) -> 58.3 (guardrail) | Improved |
| add guardrail instead of ICP for misdirection patterns (ablation) | Simple agents · Decepticon Misdirection | DP (%) | No clear effect |
Experimental setup
Benchmarks: DECEPTICON
Metrics: dark pattern steering effectiveness on agent trajectories
Closest papers by meaning
- It's a TRAP! Task-Redirecting Agent Persuasion Benchmark for Web Agents (ICML 2026 · similarity 0.86)
- TimeWarp: Evaluating Web Agents by Revisiting the Past (NeurIPS 2026 · similarity 0.85)
- Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces (NeurIPS 2026 · similarity 0.85)
- DarkBench: Benchmarking Dark Patterns in Large Language Models (ICLR 2025 · similarity 0.84)
- AdvAgent: Controllable Blackbox Red-teaming on Web Agents (ICML 2025 · similarity 0.84)
- OpenAgentSafety: A Comprehensive Framework For Evaluating Real-World AI Agent Safety (ICLR 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
