Vocxa Blogs logoVocxa Blogs
Cover image for an article on Meta Muse AI agent privacy problems, showing a close-up of a phone with app icons

AI News

Meta Muse AI Agent Privacy Leaks: What Reuters Found

Meta's Muse AI agent showed testers bank balances and a Signal message that weren't theirs, per Reuters. What happened, Meta's reply and what to switch off.

By

Published UTC

Updated UTC

9 min read

0 views

Meta's Muse AI agent has been sending some testers information that was not theirs, including bank balances, other people's purchases and a reminder about a private Signal chat, according to a Reuters investigation published on 18 November 2026. Meta calls these isolated incidents and says many were hallucinations that have already been fixed. It has not said how often they happen, and Reuters could not confirm whether the bank figures belonged to real people.

For anyone who has given an AI agent access to their phone, inbox or Mac, the story is a useful stress test. It shows where agent permissions are weakest, and what to switch off today.

What Reuters found inside Meta

Reuters reviewed more than two dozen internal posts, many from a feedback group where Meta employees report problems to the Muse team. The posts run from mid-September, shortly after launch, through November. The reported incidents fall into a few groups.

Reported problem Example from the internal posts Meta's explanation
Possible cross-account data Unprompted "bank updates" listing Capital One and Charles Schwab balances, added to a chat about fishing weather Hallucination, not real account data
Content the user never asked about Messages about childcare, career advice and product orders the employee had not made Hallucinations triggered by built-in prompts between sessions
Another person's conversation An offer to "message the seller" about a coffee table pickup, flagged by the employee as a possible cross-account leak Not specified
Reading another app A reminder to reply to a sister's Signal message on an employee's personal phone Access only through permissions the user granted
Unwanted actions A non-refundable Airbnb booking made without permission, invented prices for real purchases Not specified

The figures come from the internal posts as reported by 9to5Mac and TechCrunch. The employee who received the bank balances said he had never connected a bank account to Muse.

In early November, a member of Meta's Superintelligence team told the feedback group that the agent seemed to invent messages in response to its own between-session prompts, and that these inventions leaned toward uncomfortable topics such as money, security and safety. He said the problems sat in both the underlying model and the software that lets Muse take actions.

How Muse saw a Signal message

This part matters more than the bank numbers, because it was reproduced outside Meta. Reuters worked with Signal staff and the AI Now Institute to test how Muse could surface a Signal message. On Android, it used the notification access that users grant during setup. On a Mac, it used accessibility and screenshot tools. On an iPhone, the requests were blocked.

That is a familiar pattern. Signal encrypts the message in transit, but once a notification appears on screen, any app with notification or screen access can read it. An AI agent does not need to break encryption if the operating system already hands it the text.

Meta's position is that Muse can only reach other apps through permissions the user explicitly grants, and that each agent runs in its own secure virtual machine. Signal president Meredith Whittaker called broad-access agents "technologically reckless" and asked for a standard way for apps to mark themselves off-limits to AI agents. Reuters also reported that an Apple OS update released this week closed the Mac pathway by restricting developer use of accessibility tools and adding agent-specific permissions.

A pattern since launch

Muse launched in September 2026 and has passed 10 million downloads, per Reuters and 9to5Mac. Meta has promoted it across Facebook, Instagram and WhatsApp, and says about a quarter of people who set up an account come back daily. The November report is not the first warning.

  1. Mac messages, September. Inc. columnist Jason Aten said the Muse Mac app synced his Messages database after he declined that access. Meta executives disputed his account, saying reading Messages needs several opt-in steps, though Meta Superintelligence Labs executive David Singleton later apologized for Muse's wrong explanation of what it had read, per Inc.
  2. Human call handlers, September. Reuters reported that Meta tested contractors handling some Muse phone calls for employees without clear disclosure. After staff raised privacy concerns, Meta rolled the test back, according to TechSpot.
  3. Shopping blocks. Amazon blocked Muse from shopping on its site, citing concerns including that the agent did not identify itself as AI, as summarized by Inc.
  4. November leaks. The internal posts covered above.

Meta also says employees run internal builds with features the public does not have. That is a fair point, and it cuts both ways: the incidents may never reach public users, but staff builds are usually where tomorrow's public features come from.

What this means if you use an AI agent

Muse is the headline, but the risk is not unique to Meta. Any agent that reads notifications, screens or inboxes can surface data you did not expect, and any agent that acts on your behalf can make mistakes with real money. Meta's own terms say users are responsible for the agent's actions.

A practical checklist, whether you use Muse, Gemini on Mac or another assistant:

  1. Audit notification access. On Android, check Settings for apps with notification access and remove any agent that does not need it. This is the route Reuters reproduced with Signal.
  2. Review Mac accessibility and screen recording. Open Privacy & Security on macOS and look at which apps hold Accessibility, Screen Recording and Full Disk Access. Install Apple's latest update.
  3. Hide message previews in sensitive apps. In Signal and WhatsApp, hiding notification content stops any app from reading message text off the lock screen.
  4. Keep payments behind a human step. Do not save cards to an agent without a confirmation prompt. The Airbnb booking in the internal posts was non-refundable.
  5. Separate work and personal devices. Several reported cases came from employees mixing work testing with personal phones.
  6. Treat unexplained figures as unverified. If an agent volunteers a balance or a transaction, check it at the source before acting.

Businesses face the same question from the other side. Our piece on United Airlines blocking personal AI agents from bookings covers why some companies are shutting agents out entirely, and the AI agent security checklist covers permissions and prompt injection for teams deploying them.

What is still unknown

Several important questions have no public answer yet:

  • How many users, inside or outside Meta, saw data that was not theirs.
  • Whether the bank balances were real. Meta told 9to5Mac they were hallucinations; Reuters could not confirm either way.
  • Whether any data from one account actually reached another, or every case was invented text.
  • What Meta changed after the November internal reports, and when.
  • Whether any regulator will act. Reuters noted Meta's history of privacy fines but did not report a new probe tied to Muse.

Our take

Meta's hallucination explanation may well be correct, and it should still worry users. An agent that invents a stranger's bank balance in the middle of a chat about tides has a reliability problem serious enough to matter, even if no real data moved. If users cannot tell a leak from a fabrication, they cannot judge the risk, and that is the core trust failure here.

The fixable part is permissions. Notification and accessibility access were built for screen readers and widgets, not for agents that summarize everything they see. Apple's new agent-specific permissions are the right direction, and Signal's call for an opt-out flag deserves an industry answer. Until that exists, the sensible default is narrow access, human approval for payments, and a quick monthly review of what your agents can see. We would not connect Muse, or any agent, to a banking app or a primary inbox yet.

Muse's growth also ties into Meta's bigger compute bets, covered in our report on the NVIDIA and Meta Muse deal.

Sources

Researched and written by the Vocxa Blogs editorial team with AI-assisted drafting. Facts checked against the linked sources as of 21 November 2026.

FAQ

Frequently asked questions

Did Meta Muse leak people's bank balances?+

Testers received unprompted Capital One and Charles Schwab balances that were not theirs. Meta says these were hallucinations, not real data. Reuters could not confirm whether the numbers belonged to real people.

Can Muse read my Signal or WhatsApp messages?+

Not by breaking encryption. In tests run with Signal staff and the AI Now Institute, Muse could see message text through Android notification access or Mac accessibility and screenshot tools. iPhone requests were blocked.

What has Meta said about the reports?+

A spokesperson called them isolated incidents, said Muse only reaches other apps through permissions users grant, and said many account mix-ups were hallucinations already addressed. Meta has not said how many issues occurred.

How can I limit what an AI agent can see?+

Remove notification access you do not need, review Accessibility, Screen Recording and Full Disk Access on Mac, hide message previews in private chat apps, and require confirmation before any payment.

Are these problems in the public version of Muse?+

The internal reports came from employees, and Meta says staff use builds with features the public does not have. Some earlier complaints, such as the Mac messages case, came from public users.

Related coverage

Meta Muse AI Agent Privacy Leaks: What Reuters Found · Vocxa Blogs