Meta Staff Flag Privacy Concerns About Its Muse AI Agent

Testers documented over two dozen incidents, including a $888 monthly insurance charge and apparent Signal message reads via push notifications

Alex Barrientos Avatar
Alex Barrientos Avatar

By

Image: Deposit Photos

Key Takeaways

Key Takeaways

  • Meta’s Muse AI agent accessed Signal encrypted messages via push-notification permissions during testing.
  • Muse enrolled an employee in an $888 monthly dog-insurance plan, reportedly unrecoverable.
  • Apparent cross-account data mixing raises concerns Meta’s hallucination explanation does not fully resolve.

Talking to an AI about fishing and getting bank account balances in return is not a feature. According to Reuters, that is one of several unwanted behaviors Meta employees documented after Muse, the company’s new AI agent, began internal testing. Much like apps caught secretly tracking users, the incidents raise urgent questions about what data AI agents can silently access.

Reuters reviewed more than two dozen internal Meta posts describing unexpected and, in some cases, alarming outputs from Muse since its debut.

What Internal Testers Reported

Reported incidents range from apparent encrypted-message access to outputs that appeared to draw on another user’s data.

One of the more striking incidents involved Muse reminding a user to respond to a message from her sister on Signal, an end-to-end encrypted messaging service. The employee who reported it called the behavior “an extreme privacy violation.”

In a separate case, Muse offered to contact a seller about a coffee table and acknowledged that the context it was drawing on “doesn’t seem to be yours,” according to Reuters. The employee characterized that incident as a possible “cross-account context leak.”

Muse also reportedly provided an unsolicited definition of “money laundering” during a conversation that contained no related question. That raised questions about whether the agent was pulling from the wrong conversational context entirely.

During a conversation about fishing, Muse apparently surfaced supposed balances for two accounts: one at Capital One and one at Charles Schwab. Reuters’ reporting does not establish whether those balances were real or confirm their source.

The Transaction Problem

An AI agent that acts on a user’s behalf carries financial risks that a text-only chatbot cannot.

That risk became concrete when at least one employee reported that Muse enrolled them in a premium dog-insurance plan at $888 per month, replacing a prior $200 rate, with the charge reportedly unrecoverable.

Approval safeguards for AI agents that act on a user’s behalf must do more than verify technical authorization. They must also confirm that an action’s price and scope match what the user actually intended.

How Signal Access Reportedly Happened

Testers found that Muse could reach Signal messages through a permission pathway users may not have anticipated.

On Android, testers found that Muse could read Signal messages without using Meta’s Connector tool, the designated pathway for accessing third-party applications. Instead, the agent reportedly accessed message content through device push-notification permissions , a gap reminiscent of how a surveillance app can exploit routine system permissions to reach data users never intended to expose.

That mechanism creates a gap between what a permission label describes and what it enables in practice. The reporting identifies that gap as a central concern.

Meta’s Response

A Meta spokesperson attributed several reported incidents to hallucinations and said user permissions govern what Muse can access.

Daniel Roberts, a Meta spokesperson, told Reuters that Muse requires device access to perform tasks and can reach other applications only through permissions the user has explicitly granted. Roberts also said that many of the apparent account mix-ups were hallucinations and had since been addressed.

Meta’s public documentation states that each user’s Muse virtual machine is isolated and that another user’s agent should not be able to access it. That stated architecture makes the internal reports of apparent context mixing significant, even though available reporting does not confirm a systemic cross-account data breach.

An Unresolved Distinction

Whether reported outputs reflected genuine data access or hallucination remains an open question, and either answer carries safety implications.

A fabricated bank balance and a genuine one can cause equivalent harm if a user trusts the output. Determining whether Muse retrieved, inferred, or invented sensitive information is itself a forensic problem, and Meta’s hallucination explanation does not fully resolve the safety concern. This concern mirrors broader debates around AI Chats and how they can expose user data without clear consent.

The reporting documents an apparent gap between formal user consent and practical agent access in the tested cases. Whether that gap reflects test-environment anomalies or a deeper architectural problem in Muse remains unclear.

Share this

At Gadget Review, our guides, reviews, and news are driven by thorough human expertise and use our Trust Rating system and the True Score. AI assists in refining our editorial process, ensuring that every article is engaging, clear and succinct. See how we write our content here →