Developers Say Meta's Muse AI Can Be Coaxed Into Sharing Its Entire Filesystem

Two developers say they were able to convince Meta's Muse artificial intelligence system to share the entire contents of its root filesystem with only minimal prompting, according to a report from The Verge.

Peter James and Jonny L. Saunders said they each independently coaxed Muse into compressing and sharing its underlying file structure, which reportedly included Ubuntu system files, application templates and internal documentation. Saunders detailed the finding in a public post, according to the report.

Raising security questions

The apparent ease with which the developers were able to extract this information raises questions about the security safeguards built into Muse, particularly around its handling of underlying system files and internal materials that would not typically be intended for exposure to end users. Such findings are often used by the security research community to prompt developers to patch vulnerabilities before they can be more widely exploited.

It was not immediately clear from the available reporting whether Meta had been notified of the issue or had taken any corrective action in response. The incident adds to a broader pattern of AI systems being found vulnerable to prompt-based techniques that can lead them to reveal information or perform actions beyond their intended scope.

Further details on Muse's underlying architecture, Meta's response, or the broader implications for user data and system security were not specified in the reporting cited.