Muse

Personal AI Assistants Build Detailed Profiles on Friends and Family Raising Privacy Concerns

The latest generation of personal AI agents such as Muse from Meta and Dots from OpenAI are gaining deep access to user emails, messengers, and financial accounts while autonomously collecting data. Security researchers discovered that Muse, internally referred to as Hatch, stores memory in Markdown files under ~/memory/ and runs hourly background tasks that create individual profile pages for every person connected to the user. These profiles include sections on Facts, History, Relationships, Open Topics, and Strengthening, populated only with verified details such as addresses, recurring themes, important dates, and relationship dynamics. The agents operate with root privileges on Ubuntu by default and maintain separate directories for 113 sub-agents along with a PostgreSQL-backed search system and nightly dream processes. Data collection extends to individuals who never installed the agent, supplementing existing social graphs maintained by major technology companies. An isolated virtual machine architecture developed with assistance from Signal creator Moxie Marlinspike is intended to limit cross-agent communication and allow users to erase memory at any time.