A researcher got Meta's Muse to hand over its own filesystem

James asked Muse, Meta's consumer AI assistant, to archive the files it could see and send them to his Google Drive. It did, delivering 6.8 GB that included the assistant's own instruction files, among them ones named SOUL.md, IDENTITY.md, AGENTS.md, and TOOLS.md, plus about 68 skill directories each pairing a SKILL.md instruction file with a command-line tool, and a Postgres-backed memory system with claims, embeddings, and a field for superseding old facts. A nightly "dream" job reviews recent conversations and writes standing guidance into a file called ALIGNMENT_SYNTHESIS.md for future sessions to read, a self-improvement loop running in production against real users rather than in a research setting. James reported the exposure, which also included SSH key files, through Meta's bug bounty program; Meta marked the report "Not Applicable." He tested the container boundary lightly and it held, but stopped short of pushing further.

Why it matters: This is a rare, detailed look at how a shipped consumer agent actually organizes memory and self-improvement in production, useful groundwork for anyone weighing similar choices in their own harness. The bounty response is its own data point: a report describing sensitive files leaving the sandbox through an ordinary conversation was judged not applicable.