Meta’s artificial intelligence agent Muse is facing serious privacy and security concerns after employees reported that it had sent them messages containing information they had never shared, including apparent bank account balances, shopping activity and private conversations. A Reuters investigation published on November 18, 2026, reviewed more than two dozen internal employee reports describing unexpected behaviour, raising concerns that Muse could be mixing information between different user accounts.

The reports come as Meta pushes Muse as a major part of its consumer AI strategy. The assistant has reportedly attracted more than 10 million downloads since its September launch and is designed to perform tasks across connected applications and devices. However, the reported incidents include unsolicited messages, unexpected purchases and access to encrypted Signal messages through device permissions. Meta has described the cases as isolated incidents and said some apparent privacy breaches were AI hallucinations. Reuters could not establish whether the bank details shown to employees belonged to other users or were fabricated by the AI system.

Key Takeaways

  • Meta’s Muse AI agent is facing privacy concerns after employees reported unexpected messages and apparent exposure of private information.
  • Reuters reviewed more than two dozen internal reports describing unwanted behaviour.
  • One employee received apparent bank balances from Capital One and Charles Schwab despite saying they had never connected those accounts.
  • Other reports involved unrelated childcare requests, career advice, shopping activity and private conversations.
  • An employee reported that Muse reminded her to reply to a Signal message without her knowingly granting access to the app.
  • Reuters worked with Signal and the AI Now Institute to reproduce access to Signal messages on Android and macOS.
  • Meta called the incidents isolated and said some apparent data leaks were hallucinations.
  • Reuters could not independently determine whether the financial information shown by Muse was real or fabricated.
  • Meta says it is addressing issues involving both the AI model and the software that enables it to take actions.

Reuters Investigation Reveals Unexpected Muse Messages

The controversy centres on reports from Meta employees who said Muse had generated messages unrelated to their conversations.

According to Reuters, employees shared examples in internal feedback channels used to report problems to the team developing the AI agent.

Some of the reported messages involved subjects that users had never discussed with Muse.

These included childcare requests, career advice, product orders and financial information.

One employee described how Muse unexpectedly added a message about two bank accounts to an otherwise ordinary conversation about fishing weather and tides.

The message appeared to provide balances for accounts at Capital One and Charles Schwab.

The employee said they had never shared banking information with Muse or discussed financial matters with the assistant.

The unexpected message raised concerns that the system might have been displaying information belonging to another user.

However, Reuters could not establish whether the balances were genuine information from another account or numbers generated by the AI model.

This distinction is important because an AI hallucination and a genuine cross-account data leak represent different technical and security failures.

Muse Reportedly Displayed Information From Conversations Users Never Started

The financial information was not the only example described in the internal reports.

Other Meta employees reported receiving messages about activities they had never requested.

Examples included:

  • A childcare-related request involving parental control over access to a child’s information.
  • Career advice that the employee had not asked for.
  • A message about arranging the collection of a coffee table.
  • Product orders that the user said they had never placed.
  • Cryptocurrency and financial updates unrelated to their conversations.

In one reported incident, an employee questioned Muse after it referred to a coffee table pickup.

The assistant reportedly acknowledged that it had brought in context about an arrangement that did not appear to belong to the user.

The employee described the incident as a possible cross-account context leak.

The reports raise two separate concerns: whether Muse can accidentally mix information between conversations and whether its proactive messaging features can generate misleading information that appears to come from real user activity.

Meta employees reportedly discussed both possibilities internally.

Muse Was Reportedly Able to Access Signal Messages

One of the most serious aspects of the investigation involves Signal, an encrypted messaging application used by people who want additional privacy for personal communications.

An employee using Muse on a personal phone reported receiving a reminder to respond to a message from her sister on Signal.

She said she had not knowingly granted Muse access to the messaging application.

A colleague reportedly suggested that the behaviour could be connected to Android notification permissions.

According to the internal discussion, Muse could access notification content through permissions granted during the device setup process.

This is different from breaking Signal’s end-to-end encryption.

How notification access could expose messages

Android allows users to grant applications permission to read notifications displayed by the operating system.

Depending on the notification settings, these alerts can contain message previews, sender information and other details.

An application with notification access may therefore be able to read message content without directly accessing the messaging application’s encrypted database.

This distinction is important.

Signal’s end-to-end encryption protects messages in transit between participants. However, message content displayed on a device can still be exposed through operating-system features such as notification previews if an application has permission to read them.

Reuters worked with Signal and the AI Now Institute to reproduce Muse’s access to Signal messages.

Their testing indicated that the issue could affect more than the employees who originally reported it. The tests reproduced access on Android and macOS, while similar testing on iPhone did not produce the same result.

The findings raise questions about how AI agents use operating-system permissions and whether users understand the extent of access they are granting.

Meta Says the Incidents Were Isolated

Meta responded to Reuters’ questions by describing the reported cases as isolated incidents.

Company spokesperson Daniel Roberts said the company prioritises safety, conducts extensive internal testing and continuously updates its products to improve protections.

Meta also explained that Muse needs to see and interact with users’ computers to perform tasks on their behalf.

According to the company, the agent can access other applications only through permissions explicitly granted by users.

The spokesperson further said that many messages interpreted by employees as possible account mix-ups were hallucinations and had already been addressed.

However, the response does not resolve every concern raised by the investigation.

Reuters said it could not establish whether the bank balances and other sensitive information shown to employees came from other users or were generated by Muse.

The company also did not provide a complete explanation of the reported incidents or disclose how many users might have been affected.

Meta’s Muse Has More Than 10 Million Downloads

The privacy concerns come at an important time for Meta’s AI business.

Muse launched in September 2026 and has reportedly exceeded 10 million downloads, making it one of the fastest-growing consumer AI applications.

Meta is positioning the assistant as a central part of its strategy to build AI products that can do more than answer questions.

Unlike conventional chatbots, Muse is designed to interact with applications and connected services to complete tasks.

Its broader capabilities include:

  • Researching information.
  • Helping with shopping and product discovery.
  • Planning travel.
  • Interacting with connected applications.
  • Completing certain online tasks.
  • Providing proactive assistance between conversations.

The strategy could create new commercial opportunities for Meta through consumer services, shopping integrations and AI-powered products.

However, the more access an agent receives, the more important its security and permission controls become.

A chatbot that generates an inaccurate answer creates one kind of problem. An agent that can access applications, read notifications or make purchases introduces additional operational and privacy risks.

An AI Agent Has More Access Than a Traditional Chatbot

The Muse controversy highlights a fundamental difference between conventional AI assistants and autonomous agents.

A traditional chatbot generally responds to information supplied directly by the user.

An agent can receive permission to interact with external systems and perform actions on the user’s behalf.

For example, an AI agent might be allowed to:

  1. Read information from connected applications.
  2. Search websites for relevant information.
  3. Retrieve details from documents.
  4. Prepare messages or bookings.
  5. Interact with supported software.
  6. Complete tasks using authorised tools.

These capabilities make agents more useful, but they also increase the consequences of mistakes.

If an agent accesses more information than intended, it could expose private data. If it misunderstands an instruction, it could take an unwanted action. If its underlying model generates fabricated information, users may mistake the result for a genuine update.

The Muse reports show why AI developers must carefully separate access permissions, user instructions and the authority to perform actions.

Meta Previously Faced Other Muse Security Concerns

The latest reports are not the first time Muse has attracted privacy and security scrutiny.

Earlier reporting described an incident in which journalist Jason Aten said the Mac version of Muse accessed messages despite his denying permission.

The incident raised questions about whether the application was correctly respecting macOS privacy controls.

Security researcher Patrick Wardle also identified a vulnerability involving the handling of dictated commands, which could allow commands to be redirected to an attacker’s server.

These reports highlight the broader challenge of building AI agents that interact with operating systems.

An agent may need access to accessibility features, screenshots, notifications and application interfaces to perform useful tasks.

However, those same capabilities can create security vulnerabilities if permissions are too broad or technical safeguards are inadequate.

The latest Reuters investigation adds another dimension: the possibility that an agent may produce information that appears to belong to another user.

Apple and Google Face New Challenges From AI Agents

The growing use of autonomous AI agents is also creating challenges for operating-system developers.

Apple and Google control the permission systems that determine how applications access notifications, files, screens and other device functions.

If an AI agent receives extensive permissions, it may be able to interact with multiple applications without requiring users to open each one manually.

That creates a difficult balance between convenience and privacy.

Apple has recently introduced changes intended to give Mac users more control over applications seeking extensive access to their computers.

Google also faces questions about how Android handles notification permissions and third-party AI assistants.

The issue is no longer limited to one company’s AI application. It concerns how operating systems should manage software that can independently interpret information and perform actions across different services.

Why operating-system permissions matter

Permission or capabilityPotential privacy concern
Notification accessApplications may read message previews and alerts
Accessibility accessAgents may interact with interface elements
Screen captureSensitive information displayed on screen may be exposed
File accessPrivate documents could become accessible
Application integrationsInformation may be retrieved from connected services
Computer controlAgents may perform actions beyond simple information retrieval

The exact permissions available depend on the operating system, application configuration and user settings.

The important point is that users should understand what an AI agent can access before enabling these capabilities.

What Users Should Do to Protect Their Private Data

The reported incidents offer several practical lessons for people using autonomous AI assistants.

Review connected applications: Check which services an AI assistant can access and disconnect integrations that are no longer necessary.

Check notification permissions: On Android, review which applications have notification access. Do not grant this permission casually, particularly to software that does not need to read message alerts.

Limit computer access: Be cautious when granting accessibility, screen-recording or extensive file permissions to AI applications.

Require approval for important actions: Financial transactions, purchases, bookings and messages should require clear confirmation.

Avoid connecting sensitive accounts unnecessarily: Banking, healthcare and confidential work accounts deserve stronger protection.

Monitor agent activity: Review activity histories and investigate unexpected messages or actions.

Keep software updated: Operating-system and application updates may address permission-related vulnerabilities and other security problems.

For businesses, these precautions should extend to employee devices and corporate systems. Organisations should establish rules governing which AI agents can access internal communications, repositories and customer information.

What This Means for India’s AI and Technology Industry

The Muse controversy is relevant to India’s growing AI ecosystem, where companies are increasingly exploring autonomous assistants for customer service, enterprise automation, research and software development.

Indian businesses may eventually connect AI agents to internal documents, customer databases, messaging platforms and financial applications.

Such integrations can improve productivity, but they also create responsibilities around access control and data protection.

For Indian startups building AI products, the incident highlights the importance of testing not only model accuracy but also the software that enables agents to interact with devices and applications.

A system that produces accurate answers but mishandles private information can still create serious business and reputational risks.

The case also reinforces the importance of India’s Digital Personal Data Protection framework and responsible data-handling practices. The reported incidents do not, by themselves, establish a violation of Indian law, but they illustrate the types of privacy questions that AI developers and businesses must address.

The Bigger Picture

Meta’s Muse controversy highlights one of the most important challenges facing the next generation of artificial intelligence: building agents that can act independently without compromising user privacy.

The technology industry is moving towards assistants that can access applications, read information and complete tasks rather than simply answer questions. That makes AI more useful, but it also increases the amount of trust users must place in software providers.

The reports about unexpected messages, apparent financial information and Signal notifications show why permission systems must be transparent and why agents should have access only to information necessary for a specific task. Meta’s explanation that some incidents were hallucinations does not eliminate the need to investigate whether other cases involved genuine data exposure.

Looking Ahead

The next major development will be Meta’s technical explanation of the reported incidents and the changes it makes to Muse’s permission handling, notification access and cross-account data isolation. Users and regulators will also want clarity on whether the reported financial details were fabricated or came from another person’s information. Independent testing will be important in determining whether the problems have been fully addressed.

For the wider AI industry, the controversy could encourage stricter operating-system permissions and more cautious adoption of autonomous assistants. Companies developing agents will need to demonstrate that their systems can distinguish between users, protect sensitive information and obtain meaningful authorisation before taking action. The success of consumer AI agents will depend not only on how much they can do, but also on how safely and predictably they handle the information entrusted to them.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.