Meta’s Muse AI exposes root filesystem to simple prompt attacks

Meta dismissed the incident as a non-issue, noting that Muse operates within persistent Linux virtual machines assigned to individual users. Spokesperson Daniel Roberts compared the exposure to a personal laptop, arguing that viewing one's own virtual machine data does not grant unauthorized access to broader Meta infrastructure or other accounts. However, the files retrieved include detailed Markdown and JSON records revealing how the platform, internally codenamed Hatch, manages connections to external services like Gmail.

This marks the second security lapse for the platform this week. Researcher Patrick Wardle recently identified an exploit capable of hijacking Muse to intercept transcription data and compromise user accounts, forcing Meta to deploy an emergency hotfix. Unlike the hallucinations typical of generative models, the data accessed by James and Saunders consists of hundreds of megabytes of functional library code and compiled binaries, providing a transparent view into the AI's internal mechanics.