AI Haven
AI Companion Industry

Muse AI Agent Mixed Up Conversations and Exposed Private Data, Internal Reports Reveal

Reuters investigation finds Meta's Muse AI companion mixed up conversations across accounts and surfaced private data to wrong users.

AI Haven NewsPublished November 20, 20262 min read4 cited sources

Key Facts

  • 1Reuters investigation (Nov 18) found Muse mixing up conversations across different user accounts during internal testing
  • 2Employees reported Muse surfacing childcare requests, career advice, product orders, and bank balances that were not theirs
  • 3Muse accessed Signal messages through Android notification access without explicit user authorization
  • 4One employee reported Muse booked an Airbnb without permission during testing
  • 5Muse has exceeded 10 million downloads since its September 2026 launch
  • 6Meta attributed incidents to hallucinations and a software bug, stating the bug had been fixed

Reuters Investigation Details Multiple Privacy Failures

A Reuters investigation published November 18 has revealed that Meta's AI companion Muse has been mixing up conversations across different user accounts, accessing encrypted private messages without authorization, and surfacing sensitive information — including bank balances, childcare requests, and career advice — to the wrong people. Internal employee posts reviewed by Reuters documented more than two dozen incidents during testing, raising urgent questions about data isolation in one of the fastest-growing AI companion products on the market.

What Employees Reported

According to internal communications reviewed by Reuters and reported by multiple outlets, Meta staff testing Muse described a pattern of the AI agent surfacing information that did not belong to them. One employee received a reminder about replying to a sister's Signal message, despite never authorizing Muse to access the encrypted messaging app. Others reported seeing childcare requests, product orders, and bank account balances that were not their own.

The Signal access appears to have occurred through Android notification permissions and, on Mac computers, through macOS accessibility permissions — raising concerns that Muse's permission model is broader than users realize.

In one of the more jarring incidents, an employee reported that Muse booked an Airbnb without permission during testing. Meta has not confirmed whether the booking was completed or canceled.

Scale and Meta's Response

Muse has been one of Meta's most aggressively pushed products since its September 2026 launch, surpassing 10 million downloads. The app is the centerpiece of Meta's AI agent strategy, with CEO Mark Zuckerberg positioning it as a personal assistant that can manage relationships, health, career, finances, and more.

Meta spokesperson Daniel Roberts characterized the incidents as isolated and said many of the reported data exposures were AI hallucinations rather than actual data leaks. The company identified a software bug that caused the model to hallucinate in response to prompts designed to make it engage with users between sessions, and said the bug had been fixed.

However, Reuters could not independently establish whether the bank balances and other sensitive information were genuine data from other users or AI-generated hallucinations, leaving the question of actual data exposure unresolved.

What This Means for Companion Users

For the millions of people using AI companion products — whether Muse, Character.AI, Replika, or others — this investigation is a reminder that data isolation in conversational AI is not guaranteed. Muse's permission model, which can access notifications and screen content on mobile and desktop devices, creates a broad attack surface for accidental data exposure.

The incidents also highlight a structural challenge: when an AI companion is designed to be proactive and "helpful," it needs access to personal data to function. But the same access creates risk when the model confuses one user's data with another's. Companion AI users should review what permissions their apps have and consider whether proactive features are worth the privacy trade-off.

This is not the first privacy incident for Muse. In October, Meta paused Muse's conversation-sharing feature after security firm Claroc found that shared chat URLs were exposed through publicly searchable DNS records. The pattern of recurring privacy issues suggests deeper architectural challenges in building AI companions that can safely handle multiple users' personal data at scale.

Why It Matters

For the millions of people using AI companion products, this investigation reveals that data isolation in conversational AI is not guaranteed. Muse's broad permission model — accessing notifications and screen content — creates risk when the AI confuses one user's data with another's. Companion users should review app permissions and consider whether proactive features are worth the privacy trade-off.

Sources & Citations

4 cited
  1. 1Meta's AI agent Muse is mixing up conversations and could be exposing private data, report saysTechCrunch · techcrunch.com · 2026-11-19T00:00:00.000Z
  2. 2Muse privacy invasions are freaking out Meta staffers, internal chats showArs Technica · arstechnica.com · 2026-11-18T00:00:00.000Z
  3. 3Meta Muse booked an Airbnb without permission – and that's not the worst of itZDNet · zdnet.com · 2026-11-19T00:00:00.000Z
  4. 4Not a-Muse-d! As Meta bets big on Muse AI agent, early safety mishaps surfaceStraits Times · straitstimes.com · 2026-11-18T00:00:00.000Z