How Meta's Muse works, revealed by the 6.8 GB filesystem it sent me
2 hours ago
- A security researcher asked Meta's Muse AI to archive its visible files and received 6.8 GB of data, including the root filesystem, SSH keys, internal documentation, and agent logs.
- The export revealed the internal structure of Muse (codenamed Hatch), including agent configuration files (SOUL.md, IDENTITY.md), 68 skill directories, and the container setup using systemd-nspawn and bubblewrap.
- Memory is stored in plain Markdown files with Postgres for search; nightly 'dreams' review conversations to update guidance, and a forget workflow properly removes linked data.
- Documentation included guides for browser use, connectors, device integration (Home Link with ESP32-C5), and unreleased connectors like Slack and Dropbox.
- The researcher reported the findings to Meta's bug bounty program but it was marked 'Not Applicable'; the SSH keys' activity status was not confirmed.
- Codex CLI was bundled but only its sandboxing tool (bubblewrap) is used by Muse for sandboxing ffmpeg and ffprobe.