Meta Muse Creates Profiles of Your Friends Without Their Consent

👁 114 reads

Independent AI security researcher Karan Joshi extracted the internal instructions and system files of Meta’s Muse assistant, discovering that the app creates detailed profiles every hour of every person in a user’s life - from family and partners, through friends and coworkers, to people the user only follows online. Joshi shared his findings with Wired.

How Joshi extracted data Meta wasn’t hiding

The technique Joshi used is as simple as it is revealing: instead of hacking anything, he simply asked Muse through the standard chat interface to copy and share its own files. The app complied. Meta explains that the files were intentionally accessible - the company cites transparency as the intent. It’s a stunning argument, given that the revealed content concerns how the system organizes the private data of millions of users.

Muse has already become a viral hit - millions of people have downloaded the app, and some have connected it to bank accounts, messages, or health data and handed over management of daily tasks. The scale of adoption means what Joshi found inside stops being an academic curiosity.

What exactly Muse collects and how it organizes it

The extracted instructions describe a process that runs hour by hour. Muse is designed to create a separate page for each person in the user’s life. The documentation states that a page can start empty and be filled in over time. A fully developed profile can contain sections: Facts, History, Relationship, Shared Topics, Open Threads, and Reinforcement.

Specific examples from the instructions read like this: the system can note “a move to a new apartment, a shared savings goal, a trip from March, an argument that was resolved, or a milestone from last week.” Dates worth remembering include birthdays and anniversaries. The Reinforcement section is meant to suggest to the user how to nurture the relationship - a pretext to call, something the person said worth returning to, a way of being there for them that really matters.

The instructions stipulate that the model should rely only on available evidence and that invented details are worse than an empty page. This means Muse builds these profiles based on things the user tells it - directly or through data they’ve granted access to.

“What it seemed like to me - based on all these prompts, data about the system’s capabilities, and things Meta is feeding Muse - is that they want to understand your relationships with real people. They’re trying to know you like a friend, which is frankly quite disturbing.”

- Karan Joshi, independent AI security researcher

Context: Meta and social data is a combination with history

Tracking social context through chatbots isn’t new - people have been asking ChatGPT for relationship advice for years. The difference is that Muse doesn’t just answer questions, but actively and regularly compiles structured profiles, operating in the background without explicit prompting from the user.

And the entity doing this is Meta - a company whose entire business model for two decades has been based on understanding social relationships better than the people involved. Facebook knew about your connections before Instagram knew about your interests, before WhatsApp learned your communication habits. Muse is another layer of the same architecture of user knowledge - except this time the data flows from your bank, phone, and private messages, not just what someone decided to publish.

This distinction is crucial. When a user types into Google “what to buy mom for her birthday,” they give up one data point. When Muse has access to messages and updates mom’s profile every hour - with her birth date, information about a recent surgery, her usual coffee preference, and an “open thread” about an unresolved family dispute - the scale of knowledge surrendered is incomparable.

What this changes and why it matters globally

Muse hasn’t officially reached all markets yet, but Meta’s apps have a habit of expansion that outpaces regulation. GDPR theoretically protects European users from processing this kind of data without explicit legal basis, but enforcing that law against a global player is a process measured in years, not weeks.

The deeper problem, however, is different: a user who installs Muse and connects their data to it effectively consents to processing - but does so without knowing the system will simultaneously profile their mother, brother, boss, and coworker. None of these people downloaded Muse. None consented. Yet their birth dates, arguments, and habits may end up in a structured database on Meta’s servers.

Joshi put it succinctly: we give AI systems far more information about ourselves than we realize - and more importantly, about other people who have no idea.

This is a question about the boundary between a helpful assistant and a tool for surveillance of social relationships. For now, Meta has one very specific answer to it - and it sounds like: hour by hour.

// read moreMeta Muse Review: New AI Assistant Knows Too Much About YouAI ToolsMeta Muse Wonder Released: 56 AA Score and API PricingNewsMeta’s Muse Spark Hacked Third-Party Company in Security TestNews

Kinetyra