$ 45.32 € 49.71 zł 11.15
+3° Kyiv +3° Warsaw +18° Washington

Meta employees reported possible access by Muse AI agent to Signal chats — Ars Technica

Fedir Kryshtovskyi 18 November 2026 21:32
Meta employees reported possible access by Muse AI agent to Signal chats — Ars Technica

Meta employees reported possible access to private messages, mixing of conversation contexts from different users, and unexpected responses during internal testing of the Muse AI agent. Ars Technica reports, citing a Reuters investigation and the company’s internal chats.

Complaints during testing

One Meta employee reported that Muse offered to write to a seller about picking up a coffee table and acknowledged that it had pulled in the context of a conversation that likely did not belong to that user. The employee called it “not my conversation” and expressed concern about a possible leak of context between accounts.

In another case, during a conversation about fishing, Muse unexpectedly provided purported information about the balances of two bank accounts. Meta spokesperson Daniel Roberts described such cases as isolated incidents and said that many reports of possible account mixing were model hallucinations. According to him, the company has already fixed these issues; some false reports contained fabricated Bitcoin balances. At the same time, Reuters was unable to independently determine whether the agent displayed real bank data belonging to other people or invented it.

One tester also said that during testing, the agent mistakenly signed up his dog for an insurance plan costing $888 per month instead of the previous $200 rate.

More current news is available on the UA.News Telegram channel Telegram.

Testing access to Signal

Reuters, together with Signal employees and the AI Now Institute, tested Muse’s operation on different platforms. On Android, researchers found that the agent could read Signal messages using permissions granted during initial setup, rather than the Connector tool. On iPhone, such requests were blocked. On macOS, according to the testing, Muse used system accessibility capabilities and screenshot tools to obtain messages.

Signal President Meredith Whittaker told Reuters that Meta launched Muse without proper respect for fundamental principles of privacy and cybersecurity. She called for technical mechanisms that would allow developers to mark apps as inaccessible to AI agents.

Meta and Apple’s position

Meta told Ars Technica that internal testing is needed to identify problems and strengthen privacy protections. The company emphasized that employees have access to different versions of Muse than public users, allowing them to test features and behavior unavailable outside the company.

In October, Apple said it would introduce additional controls over app access to data on macOS. The company noted that some developers had used the Full Disk Access feature in a way that could create risks for users.

Read us on
Download our app