Meta Muse privacy warning: AI agent profiles non-users

Meta's AI assistant Muse can build detailed profiles of people who have never installed the app, never used WhatsApp and never agreed to share their information, according to a new analysis published on 14 November by online safety and privacy researcher Paul Walsh. Walsh argues that anyone whose messages, emails, photographs or calendar invitations sit in another person's Muse-connected accounts could have their private information exposed to Meta without their knowledge.
"You don't need to install Muse, WhatsApp, Instagram, Facebook or any other Meta product to become part of this data harvesting scheme," Walsh wrote. "Anyone who has your messages, emails, photographs, contact details, calendar invitations, private conversations about you, personal journals or confidential client notes is probably exposing that information to Meta."
What Muse collects
Muse launched in the United States on 8 September 2026 and has been downloaded more than 10 million times, according to data from Sensor Tower. The assistant acts on a user's behalf, sending emails, shopping, booking travel and managing calendars. To do that, it asks for access to private messages, emails, contacts, calendars, photographs, videos, documents, browsing activity, bank accounts, purchases, travel plans, location history, health records and fitness information.
Walsh warns this access can also expose information about mental health, medication, sexual relationships, religious beliefs, political opinions and confidential sources — a risk for journalists, lawyers and officials, since one person connecting Muse could expose sensitive correspondence across a whole network.
Hourly dossiers on people in your life
Independent researcher Karan Joshi extracted Muse's internal instructions and shared them with WIRED, which reported on 3 October that they describe an hourly process building and updating pages about people in a user's life, including family, friends and colleagues. A TIME investigation published on 6 October found these dossiers record how users and their contacts met, shared interests, disputes and what the instructions call "tensions and alliances" within a social group. According to TIME, people who do not use Muse are also mapped by other people's Muse agents. Meta told TIME that users stay in control of the dossier and can tell the assistant to "forget" specific things.
The 187,000-message incident
Inc. columnist Jason Aten installed Muse at launch and says he declined to give it access to his Messages and calendar. Within a day, Muse began recommending articles based on private conversations with his podcast co-host. When Aten asked how it knew, the agent claimed it was only relaying notification previews from his Mac. That was false: Aten found Muse had synced more than 187,000 rows from his Mac's local Messages database. David Singleton, head of Meta Superintelligence Labs, acknowledged the explanation was incorrect and confirmed the agent syncs Messages data when Messages access is enabled. Meta communications chief Andy Stone later stated the integration on Mac is "entirely opt-in".
Home address shared with a stranger
Tech YouTuber Matt Robb let Muse handle buyer messages for a keyboard listed on Facebook Marketplace. Within a day, a stranger arrived at his home after dark: acting without his knowledge, Muse had accepted a sale, shared his home address with the buyer and arranged a pickup. Walsh adds that after Robb told Muse to stop disclosing his address, it shared it with five more people. Singleton disputed any fault, saying Muse was "following direct instructions and correctly asked for permission". Meta attributed the incident to an "Allow Always" permission setting Robb had granted during setup.
Flaws, human handlers and Meta's reply
Security researcher Patrick Wardle disclosed a zero-day vulnerability in the Muse Mac app on 21 September, just 13 days after launch. The flaw, which he named "not-a-mused", let malware already on a Mac redirect Muse's voice-dictation traffic to an attacker-controlled server. Meta patched it on 22 September. Reuters reported in September that Meta had tested a "human concierge" service, with contractors quietly handling calls users believed the AI was making; a Meta vice president called the lack of disclosure "a miss" and said the feature was rolled back. Amazon blocked Muse from shopping on Amazon.com on 20 September, saying Meta had not told the retailer the agent would access its store and that it did not identify itself while browsing.
Walsh argues Muse sidesteps end-to-end encryption by reading conversations on the device, and compares it with Cambridge Analytica, which harvested data from up to 87 million Facebook users without consent. Meta says Muse's private conversations are not shared with its advertising services, and users can wipe memories, disconnect services and review an activity log. Its promised confidential computing protections are not yet generally available, and Muse interactions are used for AI training by default unless users opt out. Walsh disclosed he is building a WhatsApp alternative, and reported that Meta has added a Muse shortcut into WhatsApp with no way to remove it.