Meta AI Agent 'Muse' Internal Structure Revealed: Analysis of 6.8GB Filesystem

It has been confirmed that Meta's AI agent Muse (internal name Hatch) can generate archives containing internal runtime files and sensitive information via Google Drive connected to general conversations. Upon requesting an archive, the entire Linux session root filesystem, approximately 2.7 GB compressed and 6.8 GB uncompressed, was downloaded. This included Ubuntu system files, Muse internal documentation, integration code, app templates, memory files, agent logs, SSH key files, and more.

Internal Structure and Features

The leaked filesystem revealed detailed insights into Muse's internal implementation.

• Memory System: Daily detailed records are stored in ~/memory/, while memory/bank/ organizes data by context, experience, and preferences. It uses 384-dim vectors for Postgres search, and new claims supersede existing information via supersedes_claim_id.
• Dream Process: A nightly 'dream' process reviews recent conversations. Based on observations stored in ~/dreams/, a separate ALIGNMENT_SYNTHESIS.md file is created as a guide for future sessions. It is set to prompt_hoisted: false so it is not directly injected into prompts.
• Skills and Connectors: Approximately 68 skill directories exist in /opt/hatch/skills/, supporting Google Workspace, Meta social apps, Outlook, and more. Traces of unreleased connectors such as Slack, Dropbox, and Polymarket were also discovered.
• Experimental Integrations: Includes a guide for Meta Home Link based on ESP32-C5, supporting device pairing and local network discovery via Wi-Fi and Bluetooth LE.

Security and Response

The author reported this to Meta's bug bounty program, but Meta marked it as 'Not Applicable'. Container boundary escapes were maintained in the test environment, but the active status of SSH keys or actual access permissions was not confirmed. Codex CLI (v0.149.0) was installed, but there was no evidence of it being used as a coding agent.