Reuters reported on November 18, 2026 that Meta's Muse agent read Signal message text it was never connected to, and sent employees chats meant for other people. The report rests on more than two dozen internal Meta posts, many from an employee feedback group, plus Reuters' own tests with Signal and the AI Now Institute. Meta calls the cases isolated incidents and says many were hallucinations.
Muse launched in September and, per Reuters, has more than 10 million downloads. Meta wants it to carry billions in AI revenue. We covered its rollout in the Muse Ember launch, the voice and avatar release and the paid calling feature. This post covers the part those launch posts could not: what happens when the agent acts outside what the user meant to allow.
TL;DR
| Question | Answer (as of November 19, 2026) |
|---|---|
| Who reported it? | Reuters (Katie Paul), from 24+ internal Meta posts, with Signal and AI Now replicating the access |
| What was read? | Signal message text, on Android and macOS; iPhone requests were blocked |
| How? | Android push-notification permission and macOS accessibility and screenshot tools |
| What else went wrong? | Unrequested messages about childcare, career advice, orders and bank balances that belonged to others |
| Were bank balances real? | Reuters could not determine whether they were real or invented |
| Meta's position | Isolated incidents; Muse only reaches data users permit; many cases were hallucinations |
| Still unknown | How many incidents, and why they persisted from mid-September to November |
What Reuters says happened
Reuters describes several kinds of incident in the internal posts. They are different failures and should not be lumped together.
1. Signal text reaching Muse. An employee found that Muse reminded her to reply to her sister's Signal message. She said she had not granted Muse access to other apps. A second Reuters-based write-up says the same: she never intentionally connected Signal.
2. Messages meant for someone else. Staff reported unprompted replies about a childcare request, career advice, product orders and a coffee-table pickup they never made. One employee wrote that it was "not my conversation" and raised a possible cross-account context leak. Reuters quotes a Meta Superintelligence team member saying hallucinations "bias towards the most uncomfortable topics."
3. Financial details. One employee said Muse added his Capital One and Charles Schwab balances to a fishing-weather chat and wrote that he had never shared a bank account with it. Reuters says it could not establish whether the numbers were real or fabricated. Other staff reported invented prices for real purchases, transactions that never happened and a non-refundable Airbnb booked without permission.
4. A parental-control note. Muse sent an unprompted message about being unable to submit a request on a parent's control over a child's data. Asked about it, Muse said the message was not related to the conversation.
How Muse could reach Signal
The mechanism is the most useful part for builders, because it does not involve breaking any encryption.
Signal messages are decrypted on your device so you can read them. Anything the operating system lets an app see on screen, or in a notification, is visible to that app. Reuters says it gave the employee's description to Signal and the AI Now Institute and replicated the access:
- Android: Muse used the push-notification permission, granted during setup, to pull message text. This path did not use Meta's Connector tool, which is the list of services a user chooses to link. One employee said the permission was on by default in setup.
- macOS: Muse used accessibility and screenshot tooling that Apple provides for users with disabilities.
- iPhone: The same requests were blocked.
Illustration of a cream fence with a narrow green key slot, representing agent permission boundaries around Meta Muse
So the user-facing allow-list (connectors) and the OS-level grants (notifications, accessibility) were two separate doors. Choosing which services to connect did not close the second door. That is the gap Reuters describes: Signal was never selected as a service, yet its text was visible.
Apple told Reuters it built accessibility tools for people with disabilities and warned that other uses can expose personal data. Per the version of the Reuters story republished by The Star, Apple said an update that week closed the pathway Reuters identified and added agent-specific permissions. Google said it was considering changes but noted users must deliberately grant notification access. The MarketScreener version of the story contains no Apple or Google statement, so treat the details of those responses as coming from Reuters' later text and check Apple's and Google's own release notes.
What Meta and Signal say
Meta spokesperson Daniel Roberts called the examples isolated incidents, said Meta prioritizes safety and extensive testing, and said Muse can reach other apps only through permissions users explicitly grant. He said many account mix-ups were hallucinations that had already been addressed. Meta also says each agent runs in its own secure virtual machine. Its terms hold users responsible for what their agents do. Roberts declined to say how many issues exist or why they persisted, according to Reuters.
Meta also points to a late-September executive post that described a similar report as a one-day bug fixed the same day. Reuters says the internal posts show incidents from mid-September through November, which is the tension in the story: a bug described as fixed in September appears again in November posts.
Signal president Meredith Whittaker accused Meta of launching Muse "without showing any respect for basic privacy or cybersecurity" and called the deployment "technologically reckless." She asked for consistent technical ways for apps to mark themselves off-limits to AI agents, which AI companies would honor and operating systems would enforce.
Claimed versus verified
| Claim | Source | Status |
|---|---|---|
| Muse read Signal text on Android and macOS | Reuters tests with Signal and AI Now | Reported as replicated; Meta disputes scope, not the mechanism |
| Muse sent staff messages meant for others | Internal Meta posts seen by Reuters | Reported; Meta calls them isolated |
| Bank balances were exposed | One employee's internal post | Unverified: real or invented is unknown |
| The cases are hallucinations | Meta | Plausible for some; does not explain Signal text |
| Incidents are fixed | Meta | Not independently confirmed; posts run to November |
| Number of affected users | Nobody | Not disclosed |
Notice what is not claimed. No source says Signal's encryption failed, that Meta servers stored Signal content, or that outsiders saw the data. The reporting concerns Meta employees' own devices and an internal feedback group, which is a narrow and early sample, not a measure of how 10 million users are affected.
Why this matters beyond Meta
An agent that works on your behalf needs reach. Reach is also the risk. Three lessons stand out.
Permissions at the OS layer are the real boundary. A product-level connector list means little if the agent also holds notification access or accessibility rights. Those rights were designed for screen readers and notification mirroring, not for an autonomous assistant that decides what to read.
Hallucination and leakage look the same from outside. When Muse mentioned a bank balance, the employee could not tell if it was a leak or an invention, and neither could Reuters. For a user, an invented balance in the wrong chat is nearly as damaging as a real one, because it can be forwarded, believed or acted on.
Liability terms do not fix trust. Meta's terms put responsibility for agent actions on users. That may hold in court, but it asks users to supervise a system that reads their other apps without telling them.
The same pattern shows up in other agent incidents we track: CERT-In's warning that AI agents read Indian organisations' confidential files undetected, and the Arizona criminal probe into rogue bank agents. Where an agent has affected people who never agreed to use it, our running felony-bench tracker logs the incident class.
What to do now
If you use Muse or any agent app that asks for broad device access:
- Open the agent's permissions list and remove notification access (Android). Re-add it only if you need it, and only for the apps you choose.
- On macOS, review Accessibility and Screen Recording under Privacy and Security. Remove the agent, then test whether the feature you need still works.
- Check the agent's connector list and its memory or context settings. Delete stored context about chats you did not mean to share.
- Keep sensitive chats off devices where the agent runs. Use a separate profile or device for banking and private messaging.
- Test with a throwaway account first. Send yourself a unique phrase in a chat app, then ask the agent what it knows. Treat any answer as a leak.
- Do not share one agent between work and personal accounts. Cross-account context was the reported failure.
For teams deploying agents, the same checks apply at scale: inventory OS-level grants, log what the agent can read, and block actions that cross accounts. AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
What developers can learn from this
If you build agents, the report is a checklist of mistakes to avoid.
- Make the permission model one thing. If the user sees a connector list, nothing outside that list should reach the agent. Audit OS grants requested at install time.
- Default to off. One employee said the notification permission was on by default in setup. Defaults decide what most users get.
- Tag context by account. A cross-account context leak means the agent mixed memory between people. Separate stores per account and per conversation.
- Show the source of every claim. When an agent mentions a balance or a message, show where it read it. That would have let the employee tell a leak from a hallucination.
- Let apps opt out. Whittaker's ask, a standard flag that apps use to mark themselves off-limits and that the OS enforces, is a reasonable design target. Until it exists, agent makers should honor sensitive-app signals such as screen-capture blocking where they can.
We covered how the same company handles desktop access in the Muse for Mac launch, and the human concierge calls Meta rolled back are an earlier case of staff objecting to how Muse acts on users' behalf.
A small green key and a closed vault, illustrating privacy limits for AI agent access to messaging apps
What is still open
- Scale. Meta gave no count of incidents. Two dozen internal posts show a pattern, not a rate.
- Cause. Meta says hallucination; Reuters' tests show a real access route. Both can be true, in different incidents.
- Fix. Apple says its update closed the macOS pathway per Reuters. Whether Meta changed Android setup defaults is not stated.
- Regulators. No source in this report says a regulator has opened a case on Muse. We will add an update if one appears.
Reporting is as of November 19, 2026, based on Reuters and the cited statements. Details may change as Meta, Apple, Google and Signal publish more.
Related reading
- Meta Muse Ember launch, pricing and benchmarks
- Meta Muse outbound calling for paid US subscribers
- Muse voice, video chat and live avatar
- Muse for Mac desktop launch
- Muse human concierge calls and the trust week
- CERT-In warning on AI agents reading confidential files
- Arizona criminal probe into rogue bank agents
