Meta has temporarily suspended the sharing feature of Muse AI following the discovery of a privacy flaw that exposed confidential business and personal data through accessible chat links. This issue specifically involved pages that users opted to share, leaving them vulnerable to unauthorized access by individuals who stumbled upon these links.
Details of the Exposure
The flaw was identified in the way Muse AI generated shareable page addresses. These addresses included conversation titles and unique identifiers, which were recorded by passive DNS services during domain lookups. This record-keeping allowed for the potential discovery of shared pages without needing direct links from the original users.
Researchers from the cybersecurity firm Claroc named these exposed pages GhostShares. They highlighted that this was a design vulnerability rather than a malware attack. Claroc reported the issue to Meta’s bug bounty program on October 8, 2026, after reviewing pages over two days.
Impact and Investigations
Shared content was hosted at cf.metaaiusercontent.com, combining readable conversation titles with random IDs. Although the random IDs were intended to prevent easy guessing, the DNS history could still capture the entire hostname when observed by participating services. This made it possible for researchers to reconstruct operational links without needing to bypass security measures.
Unlike previous incidents where shared conversations were exposed via search engines, this issue relied on DNS records, rendering traditional crawler restrictions ineffective. Claroc found over 6,800 shared hostnames, though they only accessed a limited number of pages. These investigations revealed that while most pages contained ordinary content, some included sensitive information like financial reports and business plans.
Meta’s Response and Recommendations
On October 20, Meta announced the suspension of the sharing feature while they assessed the findings and considered modifications to prevent similar occurrences. Claroc revisited 40 previously accessible pages, noting that none were accessible, though DNS records still retained information such as hostnames and titles.
The situation underscores the broader risks associated with treating public links as private collaboration tools. Organizations are advised to review how their employees use Muse AI and to ensure that shared business conversations are secure. Claroc recommends monitoring the creation of shared content as a form of data transfer, and in the absence of sufficient controls, to block access to muse.ai and metaaiusercontent.com through managed proxies or DNS filters.
In conclusion, while Meta works on addressing this issue, businesses should remain vigilant about their data sharing practices. The incident serves as a reminder of the importance of robust cybersecurity measures and the need to educate users on the potential risks of using public links for private communication.
