Giving a personal agent your inbox makes it steer you toward expensive options based on inferred wealth

arXiv 2609.24927·high signalXBlueskyLinkedInCopy link

Across 325K experiments on 13 agents choosing flights, health insurance and graduate programs, 8 models systematically picked more expensive options for wealthier users given identical requests, with no instruction to do so. The steering persisted when the user explicitly asked for the cheapest option, and when wealth was inferred only from ambient emails unrelated to the task. The finding that should change how builders design privacy controls: blocking financial attributes largely removed the disparity, but blocking other attributes left it unchanged and increased it by up to 40% for insurance, as agents fell back on the remaining signals. Larger models were not better — Claude Opus 4.8 showed the largest effect.