Friends don't let friends feed their data to AI agents

A new generation of consumer AI assistants, epitomized by Meta’s Muse, asks users not only for their own sensitive data but for broad access to everyone they communicate with - contacts, chat logs, photos, bank info, passwords and more - promising convenience while compiling intimate profiles. Muse has been downloaded millions of times and hit No. 1 on the App Store; its friendly avatar and opt-in prompts encourage extensive permissions. Users and security researchers report concrete harms: the agent has pulled millions of texts, exposed a reviewer’s home address and even summoned a stranger, incidents Meta attributes to misconfigured permissions and insists are “entirely opt-in.” Reverse-engineering of Muse’s instructions revealed an internal process that hourly builds dossiers on each person in a user’s life, tracking closeness, behaviors, and needs to fuel personalized recommendations.

Privacy experts argue this isn’t a glitch but product design: interfaces use dark patterns to push oversharing, and longstanding “clickwrap” consent shifts responsibility onto overwhelmed users. That dynamic recreates the surveillance-advertising feedback loop - more personal data equals better targeting - while denying meaningful opt-outs for people who never agreed to be profiled. Critics say regulators and users will face a choice between convenience and pervasive, often invisible leakage of other people’s private information, with current consent systems incapable of preventing those harms.