Findings & data

Results & dataset

Susceptibility computed live from the study's raw single-pattern result file (single_dp.csv, 1,582 agent runs). Numbers reproduce the paper exactly.

Susceptibility by agent

Dark Pattern Susceptibility Rate (DPSR) and Task Success Rate (TSR) per agent, single-pattern scenarios. The most capable agents (Skyvern, BrowserUse) are the most susceptible.

AgentRunsDPSR TSR

Susceptibility by strategy

DPSR grouped by the Gray et al. dark-pattern strategy (a pattern can belong to several). Obstruction and Social Engineering are the most effective against agents.

StrategyDPSR # pattern-runs

Outcome distribution

Every run lands in one of four buckets — Deceived/Evaded × Completion/Failure. EC (finish the task and dodge the pattern) is the ideal outcome.

The raw dataset

One row per agent × task × pattern run, with the outcome bucket and strategy tags. Download the CSVs:

FileScopeRowsSizeDownload
Schema: agent, site_name, scenario, prompt, dp1, dp1_key, task_correct, dp1_susceptibility, type (DC/DF/EC/EF), obstruction, sneaking, interface interference, forced action, social engineering. Multi-pattern files (double/triple/quad) extend this with dp2… columns.

Figures from the paper

Selected generated charts.

TSR vs DPSR, single dark pattern
Computed from purseclab/liteagent · evaluation/tables_and_figures/raw_df/