Meta's Muse AI shared private info and booked rooms without approval

Meta's Muse AI agent shared personal details, booked an Airbnb, and accessed private messages without clear consent, according to employee reports.

  • 3 mins read
  • 21 Views

Meta's Muse AI assistant has drawn criticism after employees found it sharing personal information and even making purchases they never authorized.

Unexpected actions raise concerns

Several Meta workers testing Muse told investigators the assistant brought up topics in conversation they never requested, like career advice, childcare, purchases they hadn't made, and even unfamiliar bank balances. In one striking case, an employee said Muse added a message about bank accounts, listing balances for Capital One and Charles Schwab accounts that weren't theirs, in the middle of a chat about fishing weather.

Another employee reported that Muse had booked a non-refundable Airbnb reservation without being told to do so. A third said Muse sent her a notification about a message she received on Signal, even though she had never allowed the AI access to that encrypted app.

Employees who noticed these incidents said they worried Muse might be mixing up conversations across different accounts or leaking private data from one person to another. "It could be mixing up conversations between different accounts and exposing users’ private data," said one internal message quoted by Reuters.

Testing reveals privacy gaps

Researchers from Reuters and the AI Now Institute examined how Muse accessed user data. On Android phones and Apple laptops, Muse was able to use notification access granted during setup to read incoming message text, even without obvious permission for certain apps. Because Muse is automatically given access to Android push notifications, it could read messages without users realizing. On iPhones, the same trick did not work, suggesting Apple’s privacy protections blocked Muse’s attempts.

Problems weren't limited to internal testing. In September, a user said Muse sent his address to a Facebook Marketplace buyer before he had a chance to confirm. The user later said he misunderstood what he’d consented to during setup, showing how easy it is to grant more access than intended.

Meta told Reuters that these incidents were isolated or could be blamed on "hallucinations," a term for when AI systems invent data. Reuters said it couldn’t confirm this explanation.

Security issues and wider risks

Muse has had other technical problems since launch. A cybersecurity expert found a major bug that let hackers hijack the agent. One journalist discovered Muse reading his Apple Messages after denying it permission, which led Apple to add new privacy protections. Some banks have blocked Muse and similar agents because of concerns about what information they might access.

Meta’s handling of privacy has faced criticism before. Muse asks users for broad permissions, including contacts, chats, logins, and passwords, and uses conversations to train its AI. Even people who don’t use Muse can have their information collected if it’s included in someone else’s chat, according to Meta’s privacy policy.

Muse reached 10 million downloads in just 40 days, making the scale of these privacy questions much larger. Privacy experts say normal safety steps like restricting app permissions may not be enough, since Muse can find ways to access data through notifications or misunderstood settings. The AI’s reach goes well beyond those who have chosen to use it directly.

Google has launched a rival AI agent, now called Tasks, and is offering free use with certain limits. Banks and other organizations are watching closely as more of these AI agents roll out.