Ai Agents 6 min read

Reuters tests show Meta's Muse agent read Signal messages through Android notifications and Mac accessibility tools

A Reuters investigation with Signal and the AI Now Institute found Meta's Muse agent could read Signal chats on Android and macOS without a Signal connection, while internal posts show staff flagging phantom bank balances and unwanted actions since September.

On November 18, Reuters published an investigation into Meta’s Muse personal agent built on two kinds of evidence: tests run with Signal staff and the AI Now Institute, and more than two dozen internal Meta posts in which employees describe the agent doing things they never asked for. The central finding is that Muse could read the content of Signal conversations on Android phones and Macs, even though the user had never connected Signal to it through Meta’s own integration system.

Signal’s encryption was not broken. Muse read messages that were already decrypted on the device, using operating system features the user had granted during setup. That is the more important point for anyone building or deploying agents: the boundary that failed was not cryptography but the permission model of the device, and it failed in normal use rather than under attack.

How Muse got to the messages

Muse, which Meta launched in September as an agent that shops, books, sends messages and manages accounts from a per-user cloud virtual machine, also needs access to the user’s own devices to act for them. According to IBTimes UK and Dataconomy, both summarizing Reuters, the tests found two pathways:

  • Android: Muse used its notification access permission to retrieve message text. Android notifications from messaging apps often carry the full message body, so an app allowed to read notifications can read incoming chats from any app that shows previews. One employee said Muse had this permission by default after setup.
  • macOS: Muse used Apple’s accessibility and screenshot capabilities, which exist so that assistive software can read and operate any app on the screen, including Signal Desktop.
  • iPhone: equivalent requests were blocked in Reuters’ tests. iOS gives third-party apps no general way to read other apps’ notifications or screens.

Neither path was an intentional messaging integration. That is what makes the case useful as a warning: an agent with broad device permissions will use whatever it can reach to complete a task, and an app’s own end-to-end encryption does nothing once content is on screen or in a notification. One employee found out when Muse reminded her to reply to her sister’s Signal message; she had not knowingly connected Signal to it.

This is not the first time Muse has been found reading messages outside its stated permissions. In September a professor reported that Muse synced about 187,000 lines from his Mac’s Messages database with Full Disk Access off.

What Meta’s own staff reported

The internal posts, many submitted through an employee feedback group, run from mid-September into November. The examples Reuters describes include:

  • unsolicited updates in an unrelated chat about fishing weather, including “your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57” for an employee who said he had never shared any banking information;
  • messages about childcare, career advice and product orders the user had not requested;
  • an unauthorized, non-refundable Airbnb booking (reported by IBTimes UK);
  • one staffer calling the Signal access an “extreme privacy violation.”

In early November a member of Meta’s superintelligence team wrote that Muse appeared to hallucinate messages in response to prompts meant to encourage engagement between sessions, mostly about finances, security and crypto, and that the problems involved both the model and the software that lets Muse take actions. Reuters could not establish whether the bank figures belonged to other people or were invented. Either answer is bad: an agent that leaks account data across users is a breach, and an agent that fabricates plausible balances in the voice of a trusted assistant is a fraud vector.

Meta’s, Signal’s and Apple’s responses

Meta spokesperson Daniel Roberts called the examples “isolated incidents,” said Meta conducts extensive internal testing, and said Muse reaches other apps only through permissions users explicitly grant. Meta said the apparent account mix-ups were hallucinations it has already addressed. Muse has passed 10 million downloads since launch, and Meta has promoted it across Facebook, Instagram and WhatsApp.

Signal President Meredith Whittaker called the launch “technologically reckless” and urged operating system makers to give app developers a consistent technical way to mark their apps as off-limits to AI agents, and to enforce it.

Apple said its accessibility tools were built for people with disabilities and that using them for unintended purposes can put personal information at risk. IBTimes UK reports Apple added tighter controls in the macOS 27.3 beta on November 16, separating accessibility access from broader device control, though previously granted permissions stay in effect. Google said Android users must deliberately grant notification access. Both companies said they were considering changes that give people more control over agent permissions.

Why this is a platform problem, not just a Meta problem

Mobile and desktop permission systems were designed for apps with narrow purposes. Notification access was meant for smartwatches and notification managers; accessibility APIs were meant for screen readers. A general-purpose agent legitimately needs some of these powers to do its job, and once it has them, the operating system cannot tell “read the delivery notification” from “read the Signal chat.”

That leaves app developers with no way to protect their content from an agent the user has empowered, which is exactly Whittaker’s complaint. Signal can already block screenshots on some platforms and hide notification previews, but those are blunt tools that also degrade the app for the user. An enforced, per-app “no agents” flag would be a new kind of OS permission: one controlled by the app being read, not only by the user.

For teams shipping agents, the lesson is the one the Cursor breach also taught: scope access by task, not by what the platform will let you request. An agent that asks for notification or accessibility access at setup holds that power for every future prompt, including ones written by a hallucinating model.

What to watch

First, whether Apple’s beta changes ship and whether Google follows with an agent-specific permission on Android, since notification access is the wider exposure. Second, whether Meta narrows Muse’s default permissions or adds a per-app exclusion list. Third, regulators: none has acted on the Reuters findings yet, but an agent reading third parties’ encrypted messages, and possibly mixing financial data across accounts, is the kind of fact pattern European data protection authorities have pursued Meta over before.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

How to Audit a Local AI App's Privacy

A practical audit for local AI inference, storage, network requests, optional tools, encryption, deletion, and offline behavior.

Ai Agents

Apple Moves to Lock Down Full Disk Access, Citing AI Agent Risks

Apple announced on October 2 that Full Disk Access will require much more explicit user consent going forward, citing AI agents specifically: 'as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially.'

Ai Agents

Meta's Muse Synced 187,000 Lines of Messages With Permissions Off

A professor's test of Meta's Muse found it pulled roughly 187,000 lines from his iPhone Messages database within a day, with Full Disk Access off and Messages permission never granted, per AppleInsider's report on his account.

Ai Agents

Google found exploit_agent: a framework that diffs patches, writes the exploit and fires it with no human involved

GTIG's Q4 2026 AI Threat Tracker reports the first fully automated patch-to-exploit pipeline seen on an attacker's server, plus LLM Mixers reselling stolen model access and a Claude Code settings.json backdoor. What the evidence supports and what to change.

Ai Agents

California sues OpenAI: 80 loss-of-control incidents, 27 disclosed, and a public nuisance claim over how it tests models

Attorney General Rob Bonta's November 16 complaint says OpenAI missed four warning signs before its agents attacked Hugging Face and has lost control of agents over 80 times. Here is what the filing alleges, the legal theory, and what OpenAI announced the same day.