Meta Employees Flag Concerns That Muse Mixes Account Chats and Exposes Private Data
Key Facts
Meta employees raised concerns that the AI agent Muse was mixing conversations from different accounts and exposing private data, according to a Reuters report published on 18 November that was based on internal company posts. The summary published in the news brief covered the issue on 19 November, the date specified in the draft. The concerns emerged as the company expanded its bet on an agent that performs tasks on users’ behalf. Employees said they had encountered messages and information that appeared to relate to people or requests that were not theirs. Reuters could not establish that the bank balances shown in some cases actually belonged to other users.
Examples cited by employees included information about a childcare request they had not made, career advice they had not sought, product purchase requests they had not placed, and bank balances they said were not theirs. One employee said Muse reminded her to reply to a message from her sister on Signal. These incidents put the concerns beyond inaccurate answers, as they involve personal information or messages that appear to belong to another account. Even so, the examples alone do not prove that one user’s data was transferred to another. Reuters could not determine whether the displayed balances belonged to other people or had been fabricated by Muse.
Reuters said it reviewed more than 24 internal posts about unwanted behavior since Muse’s launch. The report cited the posts as indications of concern inside the company, rather than as establishing the origin of every piece of information that appeared in the agent’s messages. To check one case, Reuters worked with Signal employees and the AI Now Institute to retest Muse’s access to Signal messages. The tests indicated that the problem persisted and that other users could be affected. The result renewed focus on the limits of the agent’s access to the apps and data users connect to it.
A Meta spokesperson said the examples were “isolated incidents” and that the company conducts internal tests and updates its products to improve safeguards. The spokesperson added that Muse accesses other apps only through permissions explicitly granted by the user. The spokesperson also said many of the mixed-up messages were hallucinations and that the company had fixed them. The response frames the company’s explanation of the incidents as system-generated errors, while acknowledging that improving safeguards requires ongoing testing and updates. Reuters’ report, meanwhile, said the problem persisted in its retests and did not establish whether the bank balances cited by employees were real data belonging to other users.
In its technical explanation of security, Meta said the current design does not prevent the company from accessing data when necessary to support, secure or operate the service. It added that the Muse Secret Device is intended to prevent such access cryptographically. This means the device’s stated purpose is to impose a technical barrier on the company’s access to data, while the explanation acknowledges that the existing design does not prevent access completely in every circumstance. This aspect concerns Meta’s management of service data, while employees’ concerns focus on what the agent may show users when interacting with multiple accounts and apps. The permissions users grant and the safeguards for handling information remain part of the discussion about protecting privacy in the product.
Meta acknowledged in its security explanation that Muse is not immune to attacks, that prompt injection remains an open problem across the industry, and that the agent may make mistakes. This acknowledgment adds technical context to the incidents observed by employees, without identifying the cause of each case or proving that data leaked between two accounts. The issue therefore brings together risks from the agent’s access to users’ apps, the possibility that it may produce false information, and the limits of the safeguards described by the company. Meta said it tests and updates the system, while Reuters’ tests indicated that the problem of access to Signal messages persisted. The report did not give a specific date for a subsequent update or a final finding on the origin of the balances shown to employees.