# Meta's Muse AI Agent Mixes Up Users' Private Chats, Reports Say

- URL: https://usa-24.com/meta-muse-privacy-concerns/
- Publisher: USA 24
- Section: Artificial Intelligence (AI)
- Published: 2026-11-22T03:49:22.810Z
- Updated: 2026-11-22T03:49:22.817Z
- Topics: Meta, AI Agents, Data Privacy, Consumer Privacy, AI Hallucinations

> Reuters found Meta employees flagged Muse for mixing up conversations and surfacing data that may belong to other users. Meta says the incidents were hallucinations.

## Key takeaways

- Reuters published its investigation into Meta's Muse agent on November 18, 2026, after reviewing internal Meta posts that described mixed-up conversations.
- Reuters reports more than 10 million Muse downloads since the agent launched in the US in September 2026.
- Meta says the flagged incidents were hallucinations, not mixed-up accounts or exposed data, and that they have already been addressed.
- Reuters could not establish whether Muse showed users real bank balances or invented the figures.
- Reuters reported that Muse was set by default to access push notifications on Android phones.

Meta's Muse AI agent has been getting into people's private messages, inventing purchase details and, according to Reuters, sending some users chats that appear to belong to someone else. The reporting puts Meta Muse privacy squarely in question for a product the company is pushing hard across the United States. Meta says the incidents were hallucinations, the term for AI making things up, and that they have already been addressed.

## What Reuters found about Meta Muse privacy

Reuters published its [investigation on November 18, 2026](https://www.bnnbloomberg.ca/business/company-news/2026/11/18/as-meta-bets-big-on-muse-ai-agent-early-safety-mishaps-surface/), and TechCrunch followed on November 19, according to a [DEV Community repost of that follow-up](https://dev.to/paullovelace/metas-ai-agent-muse-is-mixing-up-conversations-and-showing-peoples-private-data-to-the-wrong-1bb7). The reporting draws on internal Meta posts describing Muse, an agent embodied by a mascot named Jolly. Employees worried that it could be mixing up conversations between different accounts and exposing users' private data. Muse launched in the US in September 2026, and Reuters reports more than 10 million downloads since then.

Reuters describes Muse as the centerpiece of Meta's plan to generate billions of dollars in AI revenue, which explains why internal warnings carry weight. Staff flagged safety problems “as recently as launch week,” and Meta proceeded with the September launch as scheduled, according to [Newscord's summary of the reporting](https://newscord.org/article/meta-employees-raise-privacy-concerns-over-muse-ai-mixing-up-accounts--Story_20261119_MetasAIagentMuseismic46c57c5).

One unnamed Meta employee wrote in an internal post that Reuters reviewed: “This seems like an extreme privacy violation.” That line is the sharpest piece of evidence in the story, and it is the one readers are most likely to remember.

## Mixed-up chats, invented purchases and a banking question

The reported incidents run from odd to alarming. One employee described Muse adding banking information to a conversation about fishing weather and tides, according to [NewsBytes](https://www.newsbytesapp.com/news/science/meta-s-muse-faces-privacy-concerns-over-leaked-user-data/story). Other examples included invented purchase information and chats that seemed meant for someone else.

Reuters could not establish whether Muse was showing users other people's bank balances or simply making up the figures. That gap matters. An invented balance is a model error that embarrasses a company, while a real one is an exposure of someone's financial data with security and legal consequences.

The number of affected users is unknown. [Pulse's summary of the story](https://pulse.adyog.com/briefs/meta-staff-say-its-muse-ai-agent-shows-users-other-people-s) lists the count as not known, and the reporting does not say how many people saw information that was not theirs.

## Why Meta calls the incidents hallucinations

A Meta representative said many of the flagged messages were hallucinations and had already been addressed, according to the [Reuters report carried by BNN Bloomberg](https://www.bnnbloomberg.ca/business/company-news/2026/11/18/as-meta-bets-big-on-muse-ai-agent-early-safety-mishaps-surface/). Meta maintains that the incidents involved model hallucinations, not mixing of accounts or disclosure of someone else's data, as [Root-Nation reported](https://root-nation.com/en/news-en/it-news-en/en-meta-s-ai-agent-muse-is-mixing-up-conversations-and-could-be-exposing-private-data/).

That explanation is plausible on its face, since language models do fabricate details. Employees, however, described specific mix-ups between accounts, which is a different kind of failure from an invented fact. The sources reviewed for this story do not show an independent audit that tests Meta's claim.

TechCrunch reported that Meta had not responded to its request for comment, according to Pulse's summary. Until Meta publishes technical detail or a third party verifies it, the hallucination explanation remains the company's version of events.

## How Muse's phone permissions could expose more than chats

Reuters reported that Muse was set by default to access push notifications on Android phones, based on a colleague's discovery described in the report, as summarized by [Rappler](https://www.rappler.com/technology/features/meta-muse-ai-agent-early-safety-mishaps/). Notifications are where banks, delivery services and messaging apps often show their most sensitive previews, so that default widens the agent's view considerably.

One employee said she never granted Muse access to another app, yet received a reminder to reply to her sister on Signal. The sources reviewed do not explain how that reminder was generated, which is the kind of gap that makes a permissions review worth doing now.

## Practical steps to limit Muse access today

You do not need to wait for Meta's investigation to tighten things up. The following steps are general guidance, not instructions from Meta:

- Open your phone's app permission settings and review what Muse can read, especially notifications and messages.
- On Android, turn off notification access for Muse if you do not need its push-based features.
- Disconnect any Messages, Signal or banking connection you did not intend to share with an AI tool.
- Keep account numbers, balances, card details and tax information out of any AI agent conversation.
- Check Meta AI's data and privacy settings. [Yahoo Tech's guidance on Meta AI privacy](https://tech.yahoo.com/cybersecurity/articles/meta-ai-warns-chatbot-conversations-072327281.html) is a reasonable place to start.

If you have already shared financial details with Muse, check your bank statements for activity you do not recognize.

## Why AI agents raise the privacy stakes

The core issue is the difference between a chatbot and an agent. A chatbot answers a question, while an agent keeps memory and connects to a person's apps, so a wrong answer can become a privacy breach. That distinction is why the Muse reporting matters beyond one product.

The same pattern is showing up elsewhere. Google now keeps [Gemini chats for two months even with history turned off](https://usa-24.com/gemini-chat-history-two-months/), and five US banks now make customers liable for losses involving outside AI agents, a [policy shift covered in a separate USA 24 report](https://usa-24.com/ai-agent-liability-bank-terms/).

A separate finding came from Dragos, which reported on November 20 that prompts sent to AI providers were logged or stolen. Dragos said the providers it tested were not named, and the finding is separate from Muse, according to its [blog post on AI supply chain security](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens).

## What to watch next

Three developments will show whether Muse's problems were a flaw or a pattern. First, watch whether Meta publishes technical detail, such as logs or an independent audit, that shows what the model actually did with user data. Second, watch for regulators or courts. No regulator, lawsuit or Federal Trade Commission action tied to Muse had been confirmed in the reporting reviewed for this story. Third, watch whether Meta changes its Android notification default or narrows what Muse can read.

The user count will matter too. With more than 10 million downloads, even a small error rate could touch a large number of accounts, and no one outside Meta yet knows how many people that is. Lawmakers are already debating how AI agents should be overseen, and the [senators' letter on AI agent cyber threats](https://usa-24.com/ai-agent-security-white-house-letter/) suggests incidents like these will shape that debate.

## Frequently asked questions

### Is Meta Muse safe to use?

Reuters reported that Muse mixed up conversations between accounts, invented purchase details and may have surfaced data belonging to other people. Meta says the incidents were hallucinations and were already addressed, but the sources reviewed do not include an independent audit. Users who want to limit risk should review app permissions and avoid sharing financial details with the agent.

### Did Meta Muse show users other people's bank balances?

Reuters could not establish whether Muse displayed real bank balances or invented the figures. One employee described banking information appearing in a conversation about fishing weather and tides. The distinction matters because real balances would be a data exposure, not just a model error.

### How many people were affected by the Muse problems?

The number of affected users is not known based on the reporting. Reuters says Muse has been downloaded more than 10 million times since its September 2026 US launch. Reuters published its investigation on November 18, 2026.

### How do I limit what Meta Muse can access?

Start by reviewing which apps, messages and notification types Muse can access in your phone's settings. On Android, turn off notification access for Muse if you do not need it. Also disconnect any messaging or banking connections you did not intend to share with the agent.

## Sources

1. [As Meta bets big on Muse AI agent, early safety mishaps surface (Reuters, via BNN Bloomberg)](https://www.bnnbloomberg.ca/business/company-news/2026/11/18/as-meta-bets-big-on-muse-ai-agent-early-safety-mishaps-surface/)
2. [As Meta bets big on Muse AI agent, early safety mishaps surface (Rappler)](https://www.rappler.com/technology/features/meta-muse-ai-agent-early-safety-mishaps/)
3. [Meta employees complained that the Muse AI agent was showing other people's data, Reuters reports (Root-Nation.com)](https://root-nation.com/en/news-en/it-news-en/en-meta-s-ai-agent-muse-is-mixing-up-conversations-and-could-be-exposing-private-data/)
4. [Meta staff say its Muse AI agent shows users other people's private data (Pulse, Adyog)](https://pulse.adyog.com/briefs/meta-staff-say-its-muse-ai-agent-shows-users-other-people-s)
5. [Meta Employees Raise Privacy Concerns Over Muse AI Mixing Up Accounts (Newscord)](https://newscord.org/article/meta-employees-raise-privacy-concerns-over-muse-ai-mixing-up-accounts--Story_20261119_MetasAIagentMuseismic46c57c5)
6. [Meta's Muse is mixing up conversations, exposing private user data (NewsBytes)](https://www.newsbytesapp.com/news/science/meta-s-muse-faces-privacy-concerns-over-leaked-user-data/story)
7. [Meta's AI Agent Muse Is Mixing Up Conversations, and Showing People's Private Data to the Wrong Users (DEV Community)](https://dev.to/paullovelace/metas-ai-agent-muse-is-mixing-up-conversations-and-showing-peoples-private-data-to-the-wrong-1bb7)
8. [AI Supply Chain Security: What Canary Tokens Found (Dragos)](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens)
9. [Meta AI privacy and Data & privacy settings guidance (Yahoo Tech)](https://tech.yahoo.com/cybersecurity/articles/meta-ai-warns-chatbot-conversations-072327281.html)
