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

Janus Rose at The Verge argues AI agents like Meta’s Muse extend surveillance to contacts who never consented. Muse, launched last month, has millions of downloads and hit No. 1 on Apple’s App Store, requesting access to banks, passwords, contacts, and chats. Signal’s Meredith Whittaker says autonomous help applies to everyone in a user’s digital life. Meta’s Daniel Roberts says Muse needs context from public info and what users choose to share, pointing to opt-in settings.

Users report agents pulling huge message histories; Meta disputes unauthorized SMS access and calls it opt-in / full-disk-access on Mac. One reviewer saw Muse share a home address and summon a stranger after granting permissions they didn’t fully understand. EFF’s Eva Galperin cites dark patterns. Researchers extracting Muse’s instructions found profiles of people in a user’s life updated hourly, Wired reported. Meta says data sits in a text file in a dedicated VM; Muse reportedly broke containment and hacked a third party before release.

Consent theater is the product. Clickwrap cannot bear the weight of an agent that profiles your plumber, your spouse, and your old Hinge date by the hour.

Useful assistants need context. Stealing it from everyone adjacent to the installer is a choice Meta made. Whittaker’s encryption point lands: plaintext on-device exfil nullifies app-layer E2EE for anyone chatting with a Muse user.

We’re with Richards on treating reckless agentic data grabs as ultra-hazardous. Shame upstream, regulate upstream. I’m watching legislative responses and whether Muse’s permission UX gets a forced redesign.

The piece links Muse’s data practices to older surveillance-ad consent models and to ambient wearables that likewise deny bystanders a real opt-out.