Meta's new general-purpose AI agent, Muse, contains a design flaw that allows researchers to easily discover and access sensitive user data shared via web links, according to a report from security firm Claroc. The vulnerability stems from how the platform constructs unique URLs for sharing content, making it possible to enumerate specific conversation links rather than relying on unguessable random strings.
Claroc stated that they successfully identified 6,800 links created by Muse users for sharing searches, conversations, and images. After opening several hundred of these accessible links, the researchers found information that was not intended for public consumption. The findings highlight a significant gap in the platform's access controls, as the standard method of using long, random text sequences to protect URLs appears ineffective in Muse's implementation.
The issue raises questions about the security posture of Meta's latest AI product. Mark Zuckerberg, Meta's chief executive officer, previously described Muse as "built from the ground up for privacy and security." However, the ability to browse user-generated content without authorization suggests that the underlying architecture does not adequately prevent unintended access to personal and business secrets shared within the application.