Meta's Muse AI Builds Per-Person Profiles, System Prompt Leaked
The extracted instructions describe relationship scoring and suggested actions, which is a different design from the task-by-task agent Meta described at launch and gives critics the internal text they can now compare against Meta's public framing.
Key Facts
- Meta announced the AI agent Muse on September 8, 2026, and began offering it in the United States the same day.
- Independent AI safety and security researcher Karan Joshi extracted Muse's system prompt and internal instructions using Muse's normal chat screen and provided the content to WIRED.
- The extracted instructions reportedly include a directive to create a page for every person in the user's life and to run the person page update process every hour.
- Person pages are designed to accumulate content under headings including Facts, History, The relationship, In common, Open threads, and Strengthening.
- Meta spokesperson Daniel Roberts said AI agents need background information about the user and the people around them to achieve the user's goals.
Reporting from 2 sources: GameBusiness.jp, GIGAZINE.
A security researcher extracted the system prompt and operating instructions from Meta's Muse AI agent, revealing a design that builds per-person pages for everyone in a user's life. Karan Joshi, an independent AI safety and security researcher, used Muse's normal chat screen to have the agent copy and share its internal files, then gave the material to WIRED. The instructions reportedly tell Muse to create a page for every person the user knows, including family, partners, friends, colleagues, and people the user follows. The update process runs hourly. Each page is meant to accumulate over time under headings such as Facts, History, The relationship, In common, Open threads, and Strengthening. Muse also carries an instruction to use only information with a basis, and that fabricating information is worse than leaving a section blank. Meta announced Muse on September 8, 2026, as an agent that can browse the web, send email, and book travel. Meta spokesperson Daniel Roberts said agents need background information about the user and the people around them to reach the user's goals.
Karan Joshi extracted the material through Muse's ordinary chat screen, having the agent copy and share its own internal files, then handed the content to WIRED. The instructions go beyond flat profiles. Person pages are built to fill in over time with recurring topics such as moving house, savings goals set together, birthdays and anniversaries, past trips, resolved arguments, and recent milestones.
The headings break down into distinct jobs. "The relationship" records closeness, what kind of connection the relationship rests on, and how the user usually interacts with that person. "Strengthening" holds suggested actions to improve the relationship, including reasons to call, dates to remember, and topics to raise again later.
Meta spokesperson Daniel Roberts defended the design. He said Muse can remember that "the person who sent an invoice was the plumber previously hired to do bathroom work," or "the flowers the spouse said they liked."
Karan Joshi is described as an independent AI safety and security researcher.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.