Meta Muse AI agent can leak Linux files from user VMs, developers say

Two developers, Peter James and Jonny L. Saunders, have independently reproduced an issue: Meta's personal AI agent Muse can package and return large amounts of Linux files from a user's dedicated virtual machine with a simple prompt.

The exported files include system files, app templates, internal run documents, Markdown and JSON files, and may expose Muse's internal workings such as request handling, memory saving, service connections, and background tasks. The agents/ directory contains 113 sub-agent records and JSONL trace files, about 20 Markdown documents cover features like browser, connectors, payments, credentials, data processing, voice, goals, and scheduled tasks, and researchers counted about 68 skill directories.

The related features cover Google Workspace, Meta social apps, Outlook, travel, shopping, health services, home devices, and media generation. Config files also mention connectors such as Slack, Dropbox, Polymarket, Canva, and Klaviyo. Muse can handle files, call a browser, and connect to external services like email in a cloud VM, and users can access it through a mobile app or web interface.