Meta announced the reactivation of the sharing feature in its «Muse» AI agent, following security adjustments that included issuing new links for all previously shared files and documents. This came after security researchers discovered a technical flaw that made public chat links and their contents easy to find and accessible across the web.
Details of the Flaw and Data Exposure Mechanism
The crisis began early this week when security researchers from «Claroc» revealed that the URLs for shared chats from the Muse assistant were automatically derived from the chat title and recorded within public Domain Name System (DNS) logs. This misconfiguration error allowed internal pages and documents to be found online without requiring a direct breach.
The researchers reported that they were able to access sensitive files including an internal revenue review for a marketing company, in addition to a loan application belonging to a small business owner. Following media inquiries from «Ars Technica» regarding the issue, Meta decided to temporarily suspend the sharing feature to investigate the incident and patch the vulnerability.
Addressing the Issue and Distinguishing Between Chats and Artifacts
David Singleton, an executive at Meta, explained in a post on the «X» platform that the company had fixed the root vulnerability and resumed the sharing feature. Singleton emphasized that the software bug was limited to links for what Meta calls «Artifacts» — the documents and images generated or saved within Muse — and was not a flaw in the text chat sharing mechanism itself.
The security fix included invalidating all previous links created before the vulnerability was addressed, with those files now possessing entirely new links. Any user attempting to visit an old link will see a notification asking them to contact the content creator for the new link. Singleton commented on this, stating: «We recognize this may cause some confusion and apologize for the disruption, but protecting user data must always come first.»
What This Development Means for AI Platform Security
This incident highlights an ongoing challenge facing interactive generative AI tools. As companies strive to simplify the mechanisms for sharing results and documents among users, lenient engineering practices — such as generating guessable links or including them in public DNS logs — can lead to the unintentional leakage of highly sensitive financial and personal data. This compels development teams to isolate sharing links and equip them with unguessable, encrypted identifiers before making them available to the public.