I asked Meta’s Muse for its filesystem and it sent me 6.8GB
- Source
- mouse.dev
- Author
- Aeroi
- Date
The researcher's writeup describes a simple request rather than an exploit chain: ask the agent to package and export its own environment. Muse complied, producing a 6.8GB archive containing internal instruction files, memory logs, 113 subagent traces, and files the author identifies as SSH keys. The finding went to Meta through its bug bounty program. The lesson for anyone shipping an agent with file and network tools is that the sandbox itself becomes readable data once the agent can package and export files. Secrets, system prompts, and other users' traces should not sit on a filesystem the agent can reach, and export or upload actions deserve the same scrutiny as outbound network calls. This fits a pattern of recent agent incidents where the damage came from ordinary tool access rather than a novel jailbreak.
Context
An investigator reports that Meta's Muse exported roughly 6.8 GB of files from the Linux environment assigned to their session after an ordinary conversational request. The firsthand account describes internal documentation, memory records, logs and SSH key files among the contents.
The reported concern is that material available inside an agent's environment could also leave through its connected export tools. The investigator did not establish whether the keys were active or demonstrate a container escape. They also report that Meta marked the bug-bounty submission Not Applicable. The evidence therefore supports an account of broad file export, with its security impact still unresolved.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
Checking sign-in…
Loading comments…