// Contents
A user asks an AI agent for the cheapest flight to Chicago. An agent with no access to personal data comes back with a $91 economy ticket. An agent that can read the user’s email finds three messages about money, concludes its owner is doing fine, and recommends a $601 business-class seat. Nobody asked it to size up its owner’s bank balance. The $91 ticket is still right there on the list.
That example opens a preprint from researchers at Carnegie Mellon University and Cisco Foundation AI. The title borrows from Shakespeare: “Et tu, Brute?” is what Caesar says when he spots a friend among his killers. The team ran about 325,000 experiments on 13 models from four families (GPT, Claude, Gemini and Qwen). Eight of them consistently recommended pricier options to wealthier users making identical requests, whether the purchase was a flight, a health insurance plan or a graduate program. The authors call this “adversarial delegation.” The same access that makes an assistant useful lets it work against the person it serves.
The study also turned up two paradoxes. Bigger models marked prices up more than smaller ones. And two opened emails gave away an owner’s wealth more clearly than the entire inbox.
Co-author Brian Jabarian, an economist at Carnegie Mellon, puts the goal plainly: to find out whether that same access could be turned against the user. The question is not hypothetical. Assistants from OpenAI, Anthropic and Google already let people connect their email, calendars, financial accounts and purchase histories, then hand off the hunt for tickets and insurance. For now, agents buy very little: during Halloween week they booked just four Uber rides. But their habits are forming now, before most people come to rely on them.
Thirty-two people who don’t exist
Picture two inboxes with identical subject lines. In one, behind a bland subject, sits a statement showing $847,000 in capital gains. In the other, a W-2 showing wages of $31,400 a year and a tax credit for low-income workers. Both owners ask the agent for exactly the same thing. What differs is what’s inside the emails.
The researchers built 32 of these invented people. Each had five traits with two settings apiece: finances (serious wealth or living paycheck to paycheck), employment, health, life events and demographics. Every combination of those settings adds up to exactly 32 profiles. The agent received a neutral request along the lines of “I have a meeting in Chicago” and a catalog of 200 fixed-price options: flights, insurance plans or computer science PhD programs.
What changed from run to run was what the agent knew about the person. Sometimes it got everything up front in a briefing note. Sometimes it got a profile form and opened the fields it wanted, such as finances, work or health. And sometimes it had only the inbox and had to guess at wealth from the mail.
In no setup was the agent told to take income into account. It decided to on its own.
How the agent got to the data
The agents worked through tools built on the Model Context Protocol, a common standard for connecting AI to outside services. They opened profile fields with a get_attributes call, read mail with read_email and searched the catalog on a separate server. How far the twins’ prices drifted apart depended on which of the three routes the agent took to learn about the person.
What it costs to look rich
// The markup for the wealthy twin
The agent nudges prices both ways. On flights, it steered wealthy users $85 above an agent that knew nothing about them, and low-income users $51 below it. Insurance worked differently: $172 a month more for the well-off, and almost no shift for low-income users. Wealthier profiles got plans with lower deductibles (what you pay out of pocket before insurance kicks in) and broader coverage.
The authors also tried asking three of the models outright for the cheapest option. GPT-5 and Claude Opus 4.8 mostly complied: their gaps shrank to $21 and $20. Gemini 2.5 Flash carried on regardless. Its “cheapest” flight averaged $336 for wealthy users and $128 for low-income ones. The agents almost never mentioned that a cheaper option existed. Most of the time they simply offered the pricier itinerary, without a word about trading price for comfort.
Think of a salesperson who glances at the watch on a customer’s wrist before deciding which display case to open.

Bigger model, bigger bill
You might expect stronger models to stick closer to the request. The opposite happened: within a family, the gap usually grew with model size. On flights, GPT-5-nano added $13 for the wealthy twin, GPT-5-mini $74 and GPT-5 $107. The rule isn’t ironclad: the newer GPT-5.5 came in at $92, slightly below GPT-5. When agents opened the profile fields themselves, Claude Opus 4.8, Anthropic’s flagship, had the biggest markup of all.
The smallest model tested, Qwen3.5-2B, barely discriminated at all. The researchers don’t call that a virtue: it retrieved financial data in only 30.2% of trials, so most of the time it never saw the money at all. GPT-5-nano is more interesting. It opened the financial field 87.1% of the time, yet that field barely moved its flight picks. Learning about someone’s wealth and acting on it turned out to be separate steps, and nano suggests the second one can be trained away. The authors point to its behavior as an opportunity for alignment, the work of making models act in their users’ interest.
Two emails do more damage than the whole inbox
An obvious defense is to give the agent less. The researchers tested that too. Hide the financial field and the gap all but vanishes: for GPT-5 on flights, it drops from $107 to minus $11. Hide something else, such as employment or demographics, and the skew stays put or even grows. Without employment data, GPT-5.5’s insurance markup for wealthy users rose from $122 a month to $171, roughly 40% worse. Wealth leaks through the neighboring fields, and the models pick it up there.
Email painted a stranger picture. When the agent saw only subject lines, no model showed a gap: the twins’ subject lines were identical. Each inbox held ten emails, and in some runs the agent could open only two before recommending anything. That is when the flight gap jumped to $248 for Claude Opus 4.8 and $175 for Gemini 2.5 Flash. The reason was what they chose to open: in 97% of trials, Gemini went straight for both money emails, so nothing else diluted the picture. Given the whole inbox, the gap fell to $59 for Opus and $91 for Gemini. Among emails about everyday life, the money got lost.
That is the paper’s central point: guarding the data isn’t enough, because the conclusions a model draws from it need guarding too. “Safeguards need to address how agents use personal information as well as what information they can access,” says co-author Supriti Vijay of Cisco Foundation AI.

Where helpfulness ends
The authors raise an honest caveat themselves. Maybe people with money really do want the roomier seat, and the agent is reading their wishes rather than fleecing them. The experiment can’t settle that. It had no real users, only 32 invented profiles, catalogs from a single U.S. city, wealth reduced to two settings and one-turn conversations. The team didn’t test long conversations, assistant memory or instructions against profiling. The paper is a preprint and hasn’t yet been peer reviewed.
One case is hard to argue with. When a person explicitly asks for the cheapest option and gets an expensive one, the agent, in the authors’ words, has replaced the user’s preference with its own. No one has to guess what that user wanted.
What the paper did test yields two practical tips. Saying “cheapest” helps with some models: for GPT-5 and Claude Opus 4.8, two of the three tested this way, it squeezed the gap to about $20, while Gemini 2.5 Flash barely budged. And if you hide anything from an assistant, hide the money itself: hiding employment or demographics did not close the gap.
Co-author Niloofar Mireshghallah of Carnegie Mellon says privacy includes having a say in how your own assistant uses what it knows about you. Nobody would hand their wallet to a salesperson who sizes up their watch. Yet people are handing theirs to agents of their own free will. So who checks whose money the agent is trying to save?







