Meta employees reported that Muse sent unsolicited messages about bank balances, purchases, childcare and private conversations, according to internal company posts reviewed by Reuters. The reports raise a serious privacy question about Meta’s AI agent, but Reuters could not establish whether the apparent financial details and other sensitive information belonged to real people or were fabricated by the model. Ars Technica’s coverage preserves that central caveat.
Reuters’ November 18, 2026, report prompted follow-up coverage from TechCrunch and Ars Technica. Meta’s response is that employee testing can involve versions unavailable to the public, and that many messages flagged as possible account mix-ups were hallucinations the company had addressed.
The documented story is therefore about employee-reported behavior, Meta’s explanation and an unresolved data-provenance question. Calling it a confirmed cross-account breach would go beyond the published evidence.
What Meta Employees Reported
Employees described Muse introducing material that did not fit their conversations or requests. According to TechCrunch’s account of Reuters’ reporting, examples included childcare requests employees had never made, unsolicited career advice, orders they had not placed and unfamiliar bank balances.
One employee discussing weather and tide information reportedly received an additional message listing a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57. The precise amounts made the response look like a financial update. Precision alone, however, cannot show that the figures came from an actual account.

Another employee received an offer to contact a seller about a coffee-table pickup unrelated to their conversation. Ars reports that the employee flagged a “possible cross-account context leak.” Muse itself reportedly suggested that it had pulled in context that did not belong to that person. An agent’s explanation of its own response is still generated output, however, not an independently verified record of where the information originated. Ars Technica
A separate report involved an employee using Muse on her personal phone. She said the agent reminded her to reply to her sister on Signal, despite her understanding that she had not granted access to other apps. TechCrunch
That last example raises a different question from cross-account exposure: whether an agent accessed the same user’s information outside the permissions they understood themselves to have granted. The published account does not establish how the reminder was produced. It also provides no basis for concluding that Muse broke Signal’s encryption.
Meta Says Many Reports Were Hallucinations
Meta told TechCrunch that employee testing is intended to identify problems so the company can improve products and implement safety and privacy protections. Its spokesperson also said employees can access different Muse versions, features and behavior from those available to the public. TechCrunch
That is relevant context. A failure in an internal build cannot automatically be assigned to the public app. But the statement does not identify which versions produced each reported incident, so it cannot establish the boundary of any public impact either.
Meta spokesperson Daniel Roberts described the reports to Reuters as “isolated incidents,” according to Ars. He also said that “many of the messages flagged by employees as potential account mix-ups were hallucinations” that Meta had already addressed. Ars Technica
The internal explanation reported by Reuters was more specific: built-in prompts encouraged Muse to engage with users between sessions, and the agent appeared to generate unsolicited messages in response. An AI team member reportedly said those hallucinations skewed toward uncomfortable subjects such as finances, security and safety. Some involved invented bitcoin balances.
That explanation offers a plausible mechanism for an unsolicited financial update without requiring access to another customer’s banking information. A prompt encouraging proactive engagement could produce a convincing-looking update with no underlying account record.
Meta’s explanation remains an attributed company account, not an independently reproduced finding. Roberts declined to answer Reuters’ questions about the prevalence or persistence of the issues discussed internally, according to Ars. And the word many leaves the disposition of other flagged messages unspecified.
Three Different Problems Need Different Evidence
The most useful way to read these reports is to separate the possible failures. Similar-looking output can result from substantially different problems.
| Possible failure | What it would mean | Evidence needed to distinguish it |
|---|---|---|
| Fabricated sensitive information | Muse generated plausible but invented balances, reminders or conversation details. | Investigation of the triggering context and relevant system records. |
| Cross-account exposure | Information associated with one person reached another person’s session. | A verified match to real source data and evidence tracing the account boundary crossing. |
| Unexpected access to the user’s own information | Muse used personal information through a route or permission the user did not expect. | The relevant permissions, enabled integrations and actual access path. |
These are analytical distinctions, not findings that any particular failure occurred. This article’s assessment rests on the published reporting, not independent testing.
A bank-like number is evidence that the agent produced a bank-like number. Establishing a leak requires connecting that output to actual information and showing how it reached the wrong recipient. Likewise, a reminder about a private message raises an access question, but its wording alone cannot reveal the technical route.
My view is that the permission question deserves attention even if the financial outputs prove entirely fabricated. Users need to understand which personal information an agent can use and when it will act on it. An unexpected reminder can reveal a mismatch between the product’s behavior and the user’s understanding without proving a cross-account breach.
Hallucination would also leave a product-safety problem. Invented financial alerts can mislead users, trigger unnecessary account checks and undermine confidence in legitimate notifications. The privacy implications differ sharply, but the output still needs an explanation.
Public Reaction Includes a Separate Login Complaint
TechCrunch and Ars Technica brought Reuters’ findings to additional readers, with Ars explicitly highlighting the unresolved distinction between genuine private data and fabricated details. Their coverage largely describes the same underlying Reuters investigation; it should not be counted as multiple independent confirmations of a leak.
Frequently Asked Questions
3 questions
1Has Meta Confirmed a Cross-Account Data Leak?
No. Meta said many flagged messages were hallucinations that it had addressed. Reuters could not establish whether the sensitive-looking output contained another person’s real information.