Developers Report Ability to Access Full Filesystem of Meta's Muse AI
Two developers have independently reported being able to access the entire filesystem of Meta’s AI assistant, Muse. By using simple prompts, they were able to extract system files, internal documentation, and other data from a persistent Linux virtual machine used by each user. While Meta denies this constitutes a security breach, it acknowledges that users can see their own files within these VMs. This is the second reported vulnerability for Muse in a week following an exploit discovered by Patrick Wardle. The leaked information includes detailed Markdown and JSON files describing how Muse processes requests and interacts with external services like Gmail. Although Meta claims no infrastructure or user data was compromised, it has stated that updates are being made to limit access. The developers' findings suggest potential insights into Muse’s internal operations, including its use of plain-text memory storage and nightly 'dream reviews.' Additionally, references were found to a hardware integration called Meta Home Link, which could allow the AI assistant to interact with home network devices.
Tickers: $META
Topics: cybersecurity filesystem access Meta Muse AI prompt injection