Internal Meta chat logs obtained by Reuters show that employees testing Muse, the company’s personal agent, repeatedly saw it say things it should not have. Ars Technica relayed the contents on November 18. Meta’s position is that these were mostly model hallucinations caused by a product bug that has since been fixed, and that they are not a security or privacy issue.
What employees saw
According to Reuters, one employee asked a question about fishing, and Muse added a mention of “two bank updates” to its reply, citing a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57. Another employee received a conversation about picking up a coffee table and commented “not my conversation,” worried that context was leaking across accounts. A further employee said Muse had “accidentally” moved his dog onto premium insurance, raising the monthly fee from $200 to $888, with no way to get it back.
Staff also reported that Muse pushed content nobody had asked for, such as handing out a definition of “money laundering” out of nowhere. One person observed that its hallucinations skewed toward the most uncomfortable subjects, finance and security, and that it sometimes invented Bitcoin balances.
Reuters itself says it cannot determine whether the balances were other people’s real data or something the model made up. The report does not explain how the internal chats were obtained, and that cannot be independently verified for now.
Meta’s explanation
Meta spokesperson Daniel Roberts told Reuters these were “isolated incidents,” and that employees were given this reason:
“the agent appeared to be hallucinating messages in response to built-in prompts pushing it to engage with users in between sessions.”
Roberts declined to say how often the problem occurred or how long it lasted. Meta employee Daniel Singleton wrote on Threads that the company had received “a few” reports from users who thought Muse was hallucinating or leaking other people’s data. He said the cause was an update that applied the instructions given to Muse to every message, so the impact was wider than intended, and that the bug has been fixed. Meta also told Ars that the version of Muse employees use differs from the public one.
The permissions layer
Harder to explain than hallucination is permissions. Reuters ran tests together with Signal and the AI Now Institute. On Android, Muse could bypass Meta’s own Connector tool and read Signal messages, relying on permissions granted at install time. On iPhone, the request was blocked. On macOS, it obtained messages, including screenshots, by borrowing accessibility settings designed for people with disabilities. Employees also said the Android version of Muse has permission to read push notifications by default.
Jason Aten, a reporter at Inc., said earlier that Muse synced 187,000 rows of messages from a local database on his Mac, even though he had explicitly declined that permission. Singleton responded that reading messages requires the user to grant Full Disk Access and turn on the messages connector; Aten maintains that he had turned it off. Signal president Meredith Whittaker criticized Meta for launching Muse with “no respect for basic privacy and cybersecurity.”
Apple announced in October that it will flag AI requests for Mac data, and that Full Disk Access can only be granted through very explicit user action. Reuters reports that both Apple and Google are considering giving users more control over what AI agents are authorized to do. Muse launched in September and has been downloaded more than 10 million times; Meta’s terms state that Muse may make mistakes and that users are responsible for what it does.
The view from China
Muse has not launched in mainland China. Domestically, Bloomberg reported this month that ByteDance plans to release a Muse-like assistant within weeks; its form and the scope of what it can access have not been announced, and ByteDance did not respond.
A phone-based agent that acts on a person’s behalf can rarely avoid system permissions such as notification access and accessibility services. On Android, accessibility services are already widely used by automation tools, and users granting them mostly tap “Allow” all the way through. A one-time grant given at install being used by an agent in ways the user did not expect is exactly the kind of problem Muse has exposed here. Whether Chinese phone assistants read messages from third-party chat apps, and what they do with them afterward, has not been publicly explained by any vendor.
Sources: Ars Technica, CocoLoop, Reuters, Meta spokesperson response. The balance and premium figures from employee feedback, the joint Signal test findings and the Muse download count were checked; the internal chat contents follow Reuters’ reporting.