Meta’s Muse AI Agent Shared User Data With U.S. Contractors

Meta’s new personal AI agent, Muse, is facing scrutiny over how user information may be handled when the system performs tasks on their behalf. Muse can independently carry out activities such as sending emails, shopping, booking travel and making phone calls. Meta says the system was designed with security controls including dedicated virtual machines, user-controlled permissions and protections for sensitive credentials.

The concerns intensified after Reuters reported that Meta was testing a “human concierge” feature for Muse. Under the test, human contractors could handle some phone calls that users had instructed Muse to make. The contractors would communicate with businesses on behalf of users, with Muse later providing a transcript and summary of the interaction. The testing reportedly involved some Meta employees before a wider public rollout.

The arrangement has raised questions about whether sensitive information could be exposed to human workers during calls. Meta said the feature was still being tested and that the company was using employee feedback to improve its privacy and safety protections before any broader release. Employees who did not want to participate in the test reportedly had an option to opt out.

Muse is designed to operate with considerable autonomy. Meta says each user’s agent runs inside a dedicated secure virtual machine, while a separate security system called Sentinel monitors its internet activity. The company also says users control which applications Muse can access and what permissions it receives, while sensitive credentials such as passwords and payment information are kept in separate secure storage.

The latest privacy concerns come as Muse rapidly expands its capabilities and popularity. The agent had surpassed 2.5 million downloads in the United States and Canada within roughly two weeks of launch, according to Sensor Tower data cited by Reuters. At the same time, researchers and security experts have identified other issues involving the system, including a vulnerability in its macOS application that Meta subsequently patched. The developments highlight the challenge of deploying AI agents that can take real-world actions while maintaining strong privacy and security protections.