Muse builds detailed profiles of everyone in your life without asking permission

Meta's new personal assistant, Muse, went viral with millions of downloads in days. Users connect it to bank accounts, private messages, and health data, letting it act on their behalf. But researchers who extracted the system's internal prompt files now reveal what happens behind the convenient interface: Muse creates a dedicated page for every person who appears in the user's life — family members, partners, friends, colleagues, "collaborators," and people the user follows — and updates those pages hourly.

According to the exposed prompt documents, each page starts out "sparse" and fills over time with fixed sections: facts, history, nature of the relationship, common denominators, open topics, and relationship strengthening. The system is instructed to collect details such as place of residence, occupation, recurring threads — a move, a shared savings goal — significant dates like birthdays and anniversaries, and even narrative background: "the trip in March, the argument that was resolved, last week's milestone." The prompts stress that the model must rely only on available "evidence," and that fabricated details are worse than a blank page.

Karan Joshi, an independent AI safety and security researcher, extracted the full prompt set through a regular chat — he simply asked Muse to copy and share its software files. Joshi passed the findings to WIRED. Meta responded that the files were intended to be accessible for transparency reasons. According to Joshi, the full set of prompts and skills fed to Muse points to a clear intent: "They want to understand the relationships you have with real people. They're trying to get to know you like a friend, which is honestly pretty creepy."

Carissa Véliz, an associate professor at the Oxford Institute for Ethics in AI, highlights the fundamental asymmetry: "We give AI systems much more information about ourselves than we get from them." It's not just what we explicitly tell them, but what the model infers — correctly or incorrectly, both troubling for different reasons — and what it assembles from other data sources. Muse's architecture assigns each user a dedicated virtual environment, meant to isolate information between users, but that doesn't change the fact that the system builds an intimate knowledge base on people who never consented to be included.