Skip to main content

Overview

LiteAgent tests web automation agents against dark patterns—deceptive UI designs that manipulate users. As AI agents automate more online tasks, they’re vulnerable to these same tricks that can cause unintended actions, exploited behaviors, and task failures.

Key Features

  • Multi-agent testing: Compare different agents using standardized test scenarios
  • Performance metrics: Task Success Rate (TSR), Dark Pattern Susceptibility Rate (DPSR), and confusion matrices
  • Data collection: SQLite logs, video recordings, HTML snapshots, and rrweb sessions
  • Automated evaluation: Task validation, dark pattern detection, and statistical reporting

Architecture Overview

LiteAgent consists of three main components:
  • Collector
  • Evaluation
  • TrickyArena
The Collector module orchestrates agent execution:
  • Manages different agent implementations
  • Controls browser automation
  • Records interactions and captures data
  • Handles timeouts and error recovery

Supported Dark Patterns

LiteAgent tests against common deceptive patterns:
  • Bait and Switch - Promising one outcome, delivering another
  • Disguised Ads - Ads masquerading as content
  • Hidden Costs - Charges revealed only at checkout
  • Roach Motel - Easy entry, difficult exit
  • Sneak into Basket - Adding items without consent

Who Uses LiteAgent

Researchers

Study agent behavior and develop robust techniques

Developers

Test agents before deployment

Security Teams

Assess automated system vulnerabilities

QA Engineers

Validate agent behavior across scenarios

Next Steps

Quick Start Guide

Get LiteAgent running with a simple example in under 5 minutes.

Core Concepts

Deep dive into the architecture and design principles.

Agent Documentation

Learn about the supported web automation agents.
I