We study how LLMs can engage in subtle deception through strategic phrasing in legislative settings, and show that re-planning and re-sampling can increase deception rates while preserving intent.
