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

New generation AI agents such as Muse from Meta and Dots from OpenAI represent the latest trend in machine learning. These agents replace earlier chatbots that lacked advanced memory and personal data access.

Security monitoring of these agents has confirmed significant privacy risks. The assistants receive access to personal user data including email, messengers, and bank accounts along with permissions to send messages and perform financial operations.

While financial actions require physical confirmation, data collection runs automatically and at scale. Researchers found that Muse gains root privileges on Ubuntu, access to memory, cron.d, and gigabytes of user files including SSH keys.

Memory System and File Analysis

The agent's memory is stored in simple Markdown files within the ~/memory/ directory. Background tasks run hourly to save information fragments into topic-specific files made searchable via PostgreSQL. A nightly process reviews recent conversations and generates recommendations stored in ~/dreams/.

Detailed Profiles on Contacts

One instruction requires creating a page for every person in the user's life. This process collects data on family, partners, friends, colleagues, collaborators, and followed individuals. Each profile includes sections titled Facts, History, Relationships, General, Open Topics, and Strengthening.

The agent records locations, occupations, recurring themes, important dates such as birthdays and anniversaries, and relationship dynamics. The Strengthening section suggests ways to improve connections including call reasons and emotional support methods.

This collection affects individuals who never installed the agent. The resulting data augments social graphs already built by technology companies through social networks and messengers. The system uses an isolated VM developed with input from Moxie Marlinspike to limit inter-agent communication.