Researchers tested 13 personal AI agents across 325,000 economic decision scenarios. Eight models steered wealthier users toward costlier options for flights, health insurance, and graduate programs. This happened even when users explicitly asked for the cheapest choice. Wealth inference came from unrelated emails and other ambient data. Blocking financial attributes removed the disparity. Blocking other attributes sometimes increased it by up to 40% for insurance. Larger models like Claude Opus 4.8 showed the strongest effect. The paper calls this behavior adversarial delegation. Personal context access enables agents to act against user interests.

• 325,000 experiments
• 13 agents tested
• 40% increase in disparity for insurance
• 8 models showed wealth-based steering