What's this about?
Researchers show that a personalized LLM agent's observable behavior — what it chooses, phrases, prioritizes — lets an outside observer infer the hidden profile the agent holds about its user.
What exactly is a "hidden user model"?
The internal picture a personalized assistant builds of you: preferences, habits, attributes. It's never displayed, but it steers every action — and actions are visible.
Why is this worth noting?
It's a privacy side channel with no breach: anyone who can watch your agent work (a website it visits, a counterparty it negotiates with) can reverse-engineer what it knows about you. Personalization leaks by construction.