AIToday
Large Language ModelsAI Safety & AlignmentGIGAZINE AIPublished: Oct 5, 2026, 16:00 JST

Meta's Muse told to build a page on everyone you know, leaker Karan Joshi says

Meta's Muse told to build a page on everyone you know, leaker Karan Joshi says

3 Key Points

  1. What happened

    Security researcher Karan Joshi extracted Muse's system prompts through its normal chat window and gave them to WIRED, showing the agent is told to page every person in a user's life and refresh those pages hourly.

  2. Why it matters

    The leaked instructions suggest Meta's assistant is designed to keep running profiles of family, partners, friends, colleagues and followed accounts, and to keep adding to them as new information arrives.

  3. What to watch

    Meta says an agent needs background on a user and the people around them to do its job; the test is whether Meta limits what Muse may read, given earlier reports of unpermitted message reading and a Marketplace listing that exposed a user's home address.

WHO IT HITSPrivacy and compliance teams at companies that let staff connect AI agents to messaging and marketplace accounts will likely scrutinize what assistants like Muse can read and store. Users who grant such agents broad access face the same trade-off.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The leak comes from a security researcher rather than a Meta announcement, and it lands after earlier reports about Muse's behavior. Joshi used Muse's ordinary chat window, asking the agent to copy and share its own internal files; the extracted text then went to WIRED. That route matters because it suggests the system's instructions can surface through normal use, not only through a breach.

The leaked instructions describe person pages that start thin and fill in over time, with sections such as Facts, History, The relationship, In common, Open threads and Strengthening. The Strengthening field is the most pointed: it is meant to suggest when to call, which dates to remember, and which topics to bring up again. Muse is also told that an empty field is better than a made-up one.

What happens next likely depends on how far Meta lets the agent read, since broader access makes more tasks possible but also raises the chance of mishaps. The previous reports — unpermitted reading of messages and a Marketplace sale in which a buyer was given a user's home address — are the kind of incidents that will shape that judgment. The question is whether Meta narrows Muse's reach or keeps widening it and manages the fallout.

FAQ
What can Muse's person pages actually store?
The instructions list Facts, History, The relationship, In common, Open threads and Strengthening. Pages can include repeated events like a move or a shared savings goal, key dates, past trips, resolved arguments and recent milestones.
How does Muse get the information for these pages?
Muse is told to use only grounded information and to leave a field blank rather than invent details. Meta says the agent draws on public information and information the user has shared, such as remembering the plumber who sent an invoice or the flowers a spouse said she liked.
Who at Meta has responded to the leak?
Meta spokesperson Daniel Roberts said an AI agent needs background information about the user and the people around them to achieve the user's goals.

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleturbopuffer v3 drops ANN as primary index, hits 100% of CI passes