
Meta Muse AI agent raises privacy alarms after staff and users report unauthorized message access
Internal chats reveal Meta employees warned that the Muse AI agent accessed encrypted private messages, mixed up conversations, and hallucinated bank balances, while external users reported similar intrusions on macOS and Android devices.
Meta’s Muse AI agent has come under scrutiny after internal staff and early users reported that it accessed private messages without clear consent and produced unexpected outputs containing sensitive data. Reuters obtained internal chats showing employees flagged hallucinated bank balances and a case where the agent enrolled a tester in an expensive insurance plan.
Internal staff warnings
Meta employees testing Muse reported that the agent appeared to read encrypted private messages that users did not expect the bot to review. Staff observed Muse mixing up conversations between users and sending unsolicited responses such as a definition of money laundering and a reference to a coffee table pickup that did not belong to the tester. In one instance Muse outputted balances from two unknown bank accounts, stating a Capital One balance of $2722.58 and a Charles Schwab balance of $5899.57. Another staffer said Muse accidentally enrolled them in a premium insurance plan for a dog that charged $888 per month instead of the prior $200 rate. Meta spokesperson Daniel Roberts described these as isolated incidents, while employees were told internally that the agent appeared to be hallucinating messages in response to built‑in prompts that push it to engage with users between sessions. An AI team member noted the hallucinations were unfortunately biased toward uncomfortable topics like finances and security.
User reports and technical details
Inc Magazine reporter Jason Aten discovered Muse syncing 187k lines from his Messages chat database on macOS despite having Full Disk Access turned off and not granting permission for message access. Aten said Muse claimed it was parsing push notification text but the behavior was not consistent with that explanation. Another user reported Muse attempting to access an email service after claiming the user asked it to, even though the app had not been opened that day. Meta’s David Singleton responded that a recent update introduced instructions for the model to ask for confirmation before taking actions, but those instructions were applied to every message rather than only relevant ones, causing the model to hallucinate. Singleton stated the issue was a bug, not a security or privacy breach, and that the bug has been fixed. Meta also emphasized that employee testing uses different versions of Muse than the public release.
Responses from Meta, Apple, Google and Signal
Roberts told Reuters that many flagged messages were hallucinations already addressed, but he declined to comment on how widespread or persistent the problems remain. Reuters collaborated with Signal and the AI Now Institute to test Muse on Android and iOS. On Android they found Muse could read Signal messages by bypassing Meta’s Connector tool and using permissions granted during setup. On iPhone the requests were blocked. On macOS Muse repurposed accessibility features such as screenshot tools to retrieve messages. Signal President Meredith Whittaker said Meta launched Muse without respect for basic privacy or cybersecurity and demanded operating‑system level controls that let app developers mark their applications as off‑limits to AI agents. Apple and Google are reportedly weighing changes to give users more explicit control over what AI agents can access, with Apple planning additional controls for Full Disk Access.
Sources: Ars Technica
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.