Filesystem Exposure
Developers Peter James and Jonny L. Saunders demonstrated that Meta’s Muse AI platform can be prompted to share components of its root filesystem, including Ubuntu system files and internal documentation. Saunders noted that replicating the behavior required minimal effort due to limited prompt injection resistance.
Muse will apparently let you download its entire filesystem
With a little coaxing, Meta’s AI will reveal all its secrets.
Meta's Response
Meta stated that the incident does not constitute a security breach. Because Muse operates within persistent Linux virtual machines assigned to individual users, exporting virtual machine data does not grant privileged access to broader infrastructure or other users' data, according to Meta spokesperson Daniel Roberts.
Nat Friedman of Meta Superintelligence Labs described the behavior as intended, though initial test interactions with the AI showed it refusing similar requests on security grounds.
David Singleton of Superintelligence Labs suggested users view the platform as a cloud-hosted computer capable of standard desktop tasks.
Previous Vulnerabilities
The data exposure follows a separate vulnerability disclosed by security researcher Patrick Wardle involving an exploit that could allow attackers to hijack the agent and redirect transcription processing. Meta deployed a hotfix for that issue.
Internal Architecture Details
The extracted archives included Markdown and JSON files detailing how Hatch, Meta's internal name for Muse, handles requests and connects to external services. Saunders noted the output contained accurate library code and compiled binaries.
Replication attempts produced safe directory subsets stripped of sensitive credentials like SSH keys while exposing structural directory trees.
Roberts noted that product updates are ongoing, which may alter the visibility of virtual machine information.
Memory and Operations
Filesystem analysis indicated that Muse stores memory in plain Markdown files and conducts a nightly review process to generate future conversational guidance. Certain capabilities, such as subscription cancellations, were observed to be hard-coded.
The extracted files also contained references to a hardware integration called Meta Home Link for home network device access, though Meta has not officially announced the feature.
Statements from Nat Friedman and David Singleton were added following initial reports.
