Artificial intelligence November 21, 2026

Meta’s Muse AI agent reportedly mixes up chats, raising privacy concerns

Meta employees testing the company’s Muse AI agent reported getting messages about childcare requests they’d never made, orders they’d never placed and bank balances they didn’t recognize. One employee asked about weather and tides and got an unsolic...

Meta’s Muse AI agent reportedly mixes up chats, raising privacy concerns

Meta’s Muse tests raise a privacy question: Whose context is the agent using?

Meta employees testing the company’s Muse AI agent reported getting messages about childcare requests they’d never made, orders they’d never placed and bank balances they didn’t recognize. One employee asked about weather and tides and got an unsolicited follow-up listing balances for Capital One and Charles Schwab. Another got a reminder to reply to her sister on Signal. She said she hadn’t granted access to other apps and didn’t remember being asked.

The accounts come from internal posts reviewed by Reuters and summarized by TechCrunch on November 19, 2026. The reports raise the possibility that Muse showed someone else’s private data, but they don’t establish where the information came from. Meta told TechCrunch that employees test versions with features and behavior that may not be available to the public.

Someone could act on a fabricated balance and lose money. If Muse pulled a real balance from another person’s account, that would be an unauthorized disclosure. Both messages could look equally credible, right down to the cents.

Where did the information come from?

The bank balances stand out because they had no obvious connection to the weather request. The model could have made up the figures, the app could have attached context to the wrong chat, or the retrieval system could have returned a record the recipient wasn’t authorized to see. Each possibility calls for a different fix. The reporting doesn’t establish that any of them happened.

When an agent presents a balance as current, its records should show whether the figure came from an authorized tool response, a saved conversation or the model itself without a supporting source. Unsolicited messages need a record of what triggered them, too. The agent might have permission to read a balance, but that alone wouldn’t explain why it sent the figure in a chat about tides.

The Signal reminder is harder to interpret. End-to-end encryption protects messages from intermediaries, but once a message is decrypted on a device, a notification preview may display some of its contents outside the app, depending on the device and its settings. There’s no evidence Muse read a notification, accessed Signal or bypassed its encryption. The employee reported what the reminder said and what she remembered authorizing. How Muse came up with it remains unknown.

Users shouldn’t have to guess why a reminder appeared. When an agent uses a connected source, it should name that source and let the user review or revoke access. If it can’t substantiate a claim about someone’s private activity, it shouldn’t state it as fact.

Permission checks at delivery

A request can start in an authenticated chat, read from a memory store, call an external service and pass work to a background job. At each step, the system needs a reliable account identity and enough information to check what it can read and where it can send the result. A conversation ID helps route a message. It doesn’t grant permission to see what’s in it.

Permissions can change while a job is queued. A user might disconnect a service before a reminder goes out, or a retry might run after the session ends. Even a result fetched with valid credentials may no longer be appropriate to reuse. Checking permissions again at delivery adds latency and another dependency. For an agent that sends sensitive updates on its own, that’s a reasonable cost.

A single-account demo won’t catch mix-ups between accounts. Test concurrent chats and retries with two synthetic accounts whose records are easy to tell apart, then verify that each response contains only data its recipient is authorized to see. Disconnect a test integration while a reminder is pending and check that the reminder never arrives. An answer-quality benchmark may miss both failures.

Debugging can create a second privacy problem. Engineers need to know which account, connector and job produced a suspect message, but copying every prompt, record and response into a widely accessible log puts the same data at risk elsewhere. Record authorization decisions, restrict access to traces and keep retention short so engineers can investigate without routinely logging raw user content.

Risks and limitations

The published examples don’t show that any public user received another person’s data, that the balances matched real accounts or what triggered the Signal reminder. They don’t establish how often the behavior occurred, either. Internal builds may use data sources or run background tasks that public builds don’t. Meta didn’t say which features were involved or whether the versions share any of the relevant components.

To determine whether a disclosure occurred, investigators would need to trace the messages to their source and establish who had permission to see any records involved. To assess the risk in the public build, they’d need to compare how the two versions retrieve, store and deliver context. An internal-only feature could still use components shared with the public build.

Muse launched in September and had passed 10 million downloads by the time TechCrunch reported on the complaints. That figure says nothing about active use, exposure or incident frequency. Teams weighing whether to connect Muse to work messages or customer records should ask which privacy controls Meta has verified in the public version.

TechCrunch’s account was published November 19, 2026, at 17:38 UTC and updated that day at 20:08 UTC. The update note says Meta’s comment was added.

Before connecting an agent to sensitive systems, ask the vendor to show how it would trace an unexpected message: where the data came from, what authorized its use and why it reached that recipient. The system’s records should answer those questions. A model’s explanation can sound plausible without telling you what happened.

Keep going from here

Useful next reads and implementation paths

If this topic connects to a real workflow, these links give you the service path, a proof point, and related articles worth reading next.

Relevant service
AI agents development

Design agentic workflows with tools, guardrails, approvals, and rollout controls.

Related proof
AI support triage automation

How AI-assisted routing cut manual support triage time by 47%.

Related article
Meta’s Muse reaches No. 2 on the U.S. App Store, but downloads lag

Meta’s new AI agent, Muse, reached No. 2 on the U.S. App Store after launching on Tuesday. That looks strong until you look at the pace. Sensor Tower says the iOS-only app has pulled in a little over 83,000 downloads in the U.S. so far, which trails ...

Related article
Meta Business Agent brings AI customer support to WhatsApp globally

Meta is rolling out its Meta Business Agent globally inside WhatsApp, moving a long-running AI customer-support test into a real product for businesses. The company said on June 3 that the agent can answer customer questions, recommend products, ...

Related article
Meta acquires Moltbook, the AI agent social network built on bot posts

Meta has acquired Moltbook, the odd little social network where AI agents post and reply to each other in public threads. Deal terms aren’t public. Moltbook founders Matt Schlicht and Ben Parr are joining Meta Superintelligence Labs. Moltbook looked ...