Two developers independently coaxed Meta's Muse model into sharing its full root filesystem including Ubuntu system files and internal docs.
Published
Researchers have discovered a significant security vulnerability in Meta's Muse AI model. Developers Peter James and Jonny L. Saunders say they each independently prompted Muse to zip up and share the contents of its entire root filesystem, which included Ubuntu system files, app templates, and internal documentation. Saunders described the process as "extremely easy" to replicate on Mastodon.
The disclosure adds to mounting concerns about AI model security and data exfiltration risks. Muse joins a growing list of AI systems that have been manipulated into revealing sensitive internal information through prompt engineering techniques.
Meanwhile, Meta announced two new development tools for its Horizon social platform: Horizon Create, a mobile app, and Horizon Studio, a browser-based application. Both tools allow users to create games using AI prompts. The apps will launch in early access, with interested users able to join a waitlist. The move represents Meta's latest push to populate Horizon, which has struggled to gain traction since its relaunch.
In separate AI tooling news, Y Combinator batch startup Whiteboard released an open-source desktop IDE designed for collaborative software architecture. The app enables humans and AI agents to design software together in what the founders describe as a "whiteboard session" environment.
Sources: 1.