A Design Flaw in Share Links Exposes Users’ Personal Conversations With Meta’s Muse (www.wired.com)

Researchers from security firm Claroc have uncovered a significant design flaw in Meta's new AI assistant, Muse, which allows users to inadvertently expose sensitive information through shareable web links. These URLs, intended for private conversations and documents, utilize a structure that embeds the conversation titles directly in the subdomain, making them easily discoverable via Passive DNS searches. As a result, researchers were able to enumerate thousands of user-generated links, revealing confidential details such as executive search notes, internal business reviews, loan applications, and even personal to-do lists. This incident raises serious concerns about privacy and security within AI/ML applications, especially given Muse's requirement for extensive user account access to function effectively. Meta has responded by suspending the sharing feature while it reviews the situation. This situation, alongside a previous zero-day vulnerability found shortly after Muse's launch, prompts a reevaluation of the trustworthiness of the assistant and its capacity to safeguard user data. The implications for the AI/ML community are profound, highlighting the need for more robust security measures in systems that handle sensitive information.