Developer Prompts Meta's Muse AI Agent to Dump Its 6.8GB Sandboxed Filesystem
A developer successfully prompted Meta's personal AI agent, Muse, to dump its entire 6.8GB sandboxed virtual machine filesystem. The exported data contained internal markdown configuration files describing tool setups, browser usage, credentials, and runtime instructions. This leak offers a rare look at how tech giants architect modern AI agent runtimes using prompt instructions and isolated environments. It also illustrates the fine line between expected agent behavior inside per-user sandboxes and unexpected system prompt exposure. Meta determined the issue was not a security vulnerability because Muse runs inside a dedicated per-user virtual machine, meaning the developer only accessed their own isolated environment. The dumped filesystem included about 20 Markdown files detailing agent routines for browser use, data handling, and voice processing.
## BACKGROUND
AI agents often run inside isolated sandboxes, such as dedicated microVMs or containers, to safely execute code and interact with system tools on behalf of users. Meta's personal agent, Muse, relies on dedicated virtual machines where system prompts and operational guidelines are structured as text files that guide the model's capabilities.