Developers say Meta’s Muse AI can be persuaded with minimal prompting to provide an archive of its virtual machine’s filesystem, including system files, application templates and internal documentation. Meta denies this constitutes a security breach, arguing that users can inspect files within their own persistent Linux virtual machines, but the disclosures may nevertheless reveal sensitive details about how Muse operates.

The files reportedly include plain-text descriptions of Muse’s request handling, data processing and connections to services such as Gmail, as well as its Markdown-based memory system and nightly conversation reviews. This is the second Muse vulnerability reported this week, following an exploit that could hijack the agent and access an account; Meta issued a hotfix for that issue and says it is continuing to adjust how much virtual-machine information users can access.