Meta's AI Assistant Muse Builds Detailed Personal Profiles for User Contacts
Extracted system prompts revealed that Meta's AI assistant Muse runs hourly background tasks to generate and update structured profile pages for every contact in a user's life. These profiles systematically track intimacy levels, interaction patterns, past shared experiences, and provide actionable recommendations for relationship maintenance. The disclosure highlights escalating privacy concerns as modern AI agents actively encourage users to link their emails, financial accounts, and calendars to harvest broad interpersonal data. While aimed at delivering highly customized assistance, this unprecedented accumulation of intimate social graph context raises major questions around long-term data retention and surveillance. AI safety researcher Karan Joshi extracted the instructions directly through Muse's standard chat interface using prompt extraction techniques. Meta mitigates security concerns by isolating each user's context inside a dedicated Virtual Machine (VM) while offering activity logs, action confirmation prompts, and options to erase stored memory.
## BACKGROUND
A system prompt is a developer-defined instruction set that establishes an AI chatbot's persona, operational rules, and behavioral boundaries. Prompt injection exploits represent cybersecurity techniques where user inputs trick large language models into overriding developer rules or exposing confidential internal instructions.