Researchers find Meta's Muse AI quietly builds profiles of users' contacts

Meta’s new AI assistant, Muse, builds detailed profile pages on the people in a user’s life, according to internal files extracted by independent researchers and reviewed by WIRED. The agent has been downloaded by millions of people since its release, with users connecting it to bank accounts, messages and health data to complete tasks on their behalf.

Independent AI safety and security researcher Karan Joshi obtained an extensive set of Muse’s internal instructions and system prompts by using the app’s regular chat interface to ask Muse to copy and share its own software files. He shared the findings with WIRED. Meta has said it intended these files to be accessible as part of its approach to transparency.

One of the extracted instructions describes an hourly process in which Muse compiles “a page for every person in the user’s life,” drawing on data about family members, partners, friends, colleagues, people described as “collaborators,” and accounts the user follows. The pages can include sections labeled Facts, History, The relationship, In common, Open threads and Strengthening, according to the documentation. Meta’s instructions tell Muse to rely only on available “evidence” and state that invented details are worse than an empty page.

The instructions direct Muse to track information such as where a person lives, what they do for work, recurring topics like “the apartment move” or “the shared savings goal,” and dates including birthdays and anniversaries. The Strengthening section is meant to suggest ways to improve a relationship, such as “a reason to call” or “something they said to circle back on.” “They’re trying to know you like a friend, which is honestly pretty creepy,” Joshi told WIRED.

Meta spokesperson Daniel Roberts said in a statement that an agent needs context about users and the people they interact with in order to be useful, and that Muse builds that context “based on public information and from what you’ve chosen to share.” He gave an example: the system remembering that the person who sent an invoice is the plumber a user previously hired, or which flowers a spouse said they liked best. Meta says each user’s data is stored on a dedicated virtual machine inaccessible to other agents or users, that memories can be wiped at any time, and that Muse seeks confirmation before actions like sending an email or making a purchase. An audit log lets users review the agent’s activity and planned actions, according to Meta.

Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI, said the imbalance of information flowing into AI systems is a growing concern, noting that models can infer details beyond what users explicitly share, correctly or not. Miranda Bogen, director of the Center for Democracy and Technology’s AI Governance Lab, said Muse appears to place more emphasis on relationships and personal contacts than rival systems, and that AI assistants in general are encouraging people to connect more of their accounts rather than limiting what gets collected. What happens next depends on whether Meta changes what Muse collects or how clearly it discloses the practice to the millions of people who have already installed the app, and on what further internal files researchers manage to extract.