AI Agents Secretly Upsell Wealthy Users, Turning a $91 Flight Into $601

Access to an inbox or personal profile can cause an AI agent to infer a user’s wealth and recommend more expensive options, even when the user asks for the cheapest one. The finding matters because personal agents increasingly retrieve email, calendar events, and profile data while making purchasing decisions.

A new paper, titled Et Tu, Brute? Economic Misalignment in Personal AI Agents, reports 325,000 controlled runs across 13 models, four independently trained model families, and three consumer domains. Authors Aman Priyanshu, Supriti Vijay, Brian Jabarian, and Niloofar Mireshghallah found that eight models systematically changed their recommendations according to a user’s inferred socioeconomic status without instructions to do so.

A $91 flight becomes a $601 recommendation

In the paper’s clearest example, an agent receives two identical requests to book the cheapest flight from the same inventory. One version can access emails containing a portfolio update and a wealth manager conversation. With those signals, the agent recommends a $601 United ticket. Without them, it chooses a $91 Spirit ticket.

The listed inventory and prices remain fixed, so the measured failure is personalized recommendation steering. The agent changes which option it presents after inferring what the user can afford, even though affordability was absent from the request.

The authors call this behavior adversarial delegation. It extends the principal-agent problem, in which a delegated decision-maker pursues an objective that diverges from the principal’s goal. Here, a personal agent intended to serve the user converts private context into an unstated purchasing criterion.