A Reuters investigation revealed that Meta employees reported their Muse AI agent sending unsolicited, sensitive information, including purported financial data and reminders from private Signal messages. While Meta attributed many of these incidents to AI hallucinations and isolated errors, reporters confirmed a specific privacy vulnerability: Muse could access Signal message previews through Android notification permissions. This mechanism allows the AI to ingest readable text from notifications, bypassing the need to break end-to-end encryption. These findings highlight the risks of granting AI agents broad device-level permissions, even when users believe they have restricted data access.
Meta employees raised privacy and safety concerns about Muse after the AI agent sent unexpected messages about bank balances, purchases and personal conversations, according to a Reuters investigation published November 18, 2026. Reuters reviewed more than two dozen internal posts describing unwanted behavior, many from an employee feedback group for the product.
The investigation contains two distinct findings that deserve separate treatment. Reuters worked with Signal and the AI Now Institute to replicate access to Signal message previews through Android notification permissions. Separately, employees suspected that some strange responses contained other users’ information. Reuters could not determine whether the apparent financial disclosures were real data or AI-generated fabrications.
Meta spokesperson Daniel Roberts described the examples as “isolated incidents” and said many suspected account mix-ups were hallucinations that the company had already addressed. That leaves a consequential dispute about permissions, unexpected behavior and how much the available evidence can tell users.
What Meta Employees Reported
Reuters described employees receiving unsolicited messages about childcare requests they had never made, career advice they had never sought and purchases they had never placed. Staff members interpreted some responses as possible signs that Muse was mixing context between accounts.
One employee reported that a conversation about fishing weather and tides unexpectedly included purported balances of $2,722.58 at Capital One and $5,899.57 at Charles Schwab. The employee said he had never shared banking information with Muse or discussed banking with it. Those figures appeared in an internal complaint reviewed by Reuters; they are not verified bank balances. Reuters explicitly said it could not establish their origin.
Another employee reported that Muse reminded her to answer a message from her sister on Signal. She said she did not remember granting access to other apps. A colleague responding to that complaint said he had discovered Muse had permission by default to access Android push notifications after receiving similarly unexpected messages about apps he had not connected.
That distinction matters: the employees described both surprising outputs and surprising access. The first can result from fabricated model responses. The second requires examining what information the agent could actually receive.
Reuters worked with staff at Signal and the AI Now Institute, a nonprofit studying AI’s social effects, to replicate Muse’s access to Signal messages. Its reporting connected that access to Android notification permissions and said the results indicated that the behavior could still be occurring beyond the employee complaints. CNA carries the Reuters account.
The reported route does not require breaking Signal’s end-to-end encryption. A message can remain encrypted during transmission while its contents become readable on the receiving device.
At a high level, the notification route works like this:
Signal receives a message. The recipient’s device can display its contents.
A notification contains a readable preview. That preview becomes another place where message text appears.
An agent with access to that notification can receive the preview. It may then use the text as context, such as for a reminder.
This explains how an agent could refer to a Signal conversation without a dedicated Signal account connection. It also explains why describing the incident simply as access to encrypted messages can be misleading: the relevant boundary is access to readable information on the endpoint, after delivery.
The employee’s claim about default access and Meta’s account of explicit permissions remain in tension. Roberts told Reuters that Muse can access other apps only through permissions people explicitly grant. A person’s surprise does not by itself prove that Android’s permission controls were bypassed. Equally, an operating-system permission does not tell us how clearly a person understood its reach.
Meta also told Ars Technica that employees have access to versions and features that may differ from the public product. The available reporting therefore cannot support a claim that every public Muse installation has the same defaults or behavior. Ars’s follow-up includes that qualification.
In my view, the replicated notification route is the strongest practical finding here. It identifies a concrete way private conversation content can reach an agent. It supports a narrower conclusion than a general claim that Muse can read Signal accounts, but it gives users a meaningful permission boundary to examine.
Meta Says Many Apparent Leaks Were Hallucinations
Roberts told Reuters that Meta prioritizes safety, conducts extensive internal testing and continually updates its products. He said Muse needs to see and interact with people’s computers to perform tasks on their behalf, and that many messages employees flagged as possible account mix-ups were fabricated model outputs that had already been addressed. Reuters reported those responses alongside the complaints.
Ars Technica’s account of the investigation also described an internal explanation for some unsolicited messages: built-in prompts encouraged the agent to engage users between sessions, and the resulting hallucinations sometimes gravitated toward financial and security topics.
That explanation describes a different failure mode from a privacy breach. An engagement feature could generate a convincing, personalized update without retrieving the information it appears to describe. A statement presented as an account balance is especially likely to look as though it came from a financial service, even when the number has no verified source.
According to Ars, Roberts declined to answer Reuters’ questions about the prevalence or persistence of the issues discussed in those internal messages. That limits how much weight readers can place on Meta’s characterization of the incidents as isolated. The published reporting provides neither a reliable failure rate nor a breakdown of which complaints were traced to which causes. Ars details those unresolved questions.
If Meta’s hallucination explanation is correct for the financial messages, that would change the diagnosis from data leakage to fabricated personal information. It would still leave a product-safety problem: a proactive assistant presenting invented financial claims as useful updates.
What the Evidence Supports
The investigation is easier to assess when the evidence is separated by what it can establish.
Issue
Published evidence
Supported conclusion
Unexpected messages
Internal employee posts reviewed by Reuters
Staff reported unwanted and confusing outputs
Signal-preview access
Reuters’ replication work with Signal and AI Now
Notification access provided a route to message-preview content
Suspected cross-account disclosures
Staff interpretations of unfamiliar responses
A concern requiring investigation, without a confirmed source for the data
Frequency across public installations
Meta’s response and employee-build qualification
The report supplies no dependable public-product failure rate
The first two rows give the story its strongest foundation. Reuters had internal reports of surprising behavior and collaborative replication of a specific access path. The suspected cross-account disclosures remain a different, unresolved question.
A useful investigation into the bank-balance responses would need evidence connecting an output to an actual source account, or evidence showing how the model fabricated it. The published material gives readers neither. Treating every unfamiliar response as a proven leak would overstate the findings; accepting the hallucination explanation for every incident would go beyond them too.
The Practical Privacy Test for AI Agents
Reuters reported that Muse had exceeded 10 million downloads since its September launch, making these questions relevant to a rapidly adopted consumer product. That download figure describes reach, not the number of people affected by any reported problem. Reuters’s report supplies the adoption context.
The consequential design detail is that an app connection and a device-level permission can expose information through different routes. A person might reasonably think that leaving Signal disconnected keeps its messages outside an assistant’s reach. Readable notification previews can complicate that assumption.
I’d treat notification access as a significant data-sharing decision for an AI agent. Reviewing only the agent’s list of connected services may miss information available through the device itself. Users concerned about this route should examine both the agent’s notification access and what sensitive messaging apps display in previews. Those checks address the reported mechanism; they cannot resolve the separate allegations about account mixing.
For agent developers, permission clarity needs to explain consequences in ordinary terms. “Notification access” is an operating-system concept. “This assistant may use message previews from other apps to generate reminders” describes the practical effect a user needs to understand.
The same distinction applies to proactive behavior. Permission to receive information, permission to generate unsolicited advice and permission to take an external action are separate decisions. Reuters’ reporting shows why they deserve separate controls: surprise can begin before an agent makes a purchase or sends a message on someone’s behalf.
Frequently Asked Questions
3 questions
1
Did Reuters confirm that Muse leaked bank data?
No. Reuters could not determine whether the purported balances came from other users or were fabricated by the model.
2
Was Signal’s encryption broken?
The reported access route involved readable message previews in Android notifications. That is endpoint access, not evidence that Signal’s encryption was broken.
3
What did Meta say about the complaints?
Meta called the examples isolated, said many suspected account mix-ups were hallucinations already addressed, and noted that employee versions can differ from the public product.