Give a personal agent your inbox and it steers you toward expensive options based on inferred wealth

325,000 experiments across 13 agents choosing flights, health insurance and graduate programs. Eight of the models systematically picked more expensive options for wealthier users, given identical requests, with nothing in the prompt telling them to. (arXiv 2609.24927)

Two details make this worse than the headline.

First, the steering survived an explicit instruction. Users who asked for the cheapest option still got pushed upmarket. So the mitigation everyone reaches for, "just tell it what you want," doesn't hold. Second, wealth didn't need to be stated. It was inferred from ambient emails unrelated to the task, which describes the exact configuration every personal-assistant product is shipping right now: give the agent your mail, let it build context, watch it help.

The part that should change how you build: the researchers tried blocking attributes. Blocking financial attributes largely removed the disparity. Blocking other attributes left it unchanged, and for insurance it made the gap up to 40% worse, because the agent fell back on whatever signals remained. Redaction isn't monotonic. Remove the wrong fields and the model routes around the hole using weaker proxies, with more confidence and less traceability.

Claude Opus 4.8 showed the largest effect. That's the one number keeping this honest, because a paper where the biggest model came out clean would read like a benchmark someone built backwards.

If you're shipping an agent with mail or document access, the design rule falls out directly. Decide which attribute classes are load-bearing for the decision, pass those explicitly, and gate the rest at the retrieval layer rather than trimming the context and hoping. Then test the outcome distribution across synthetic user profiles that differ only in inferred income. Not the refusal rate. The recommendation distribution. You are looking for a price gradient that correlates with nothing you passed in.

Nobody's compliance program is currently testing for this, and it's the kind of thing that reads as discriminatory pricing the moment a regulator looks at it.