Claroc 揭示 Grok Build 共享連結漏洞,企業資料外露
Vibe Coding Boom Exposes Corporate Data as AI App Builders Face Security Gaps
安全公司 Claroc 發現,Grok Build 的應用共享連結可經被動 DNS 大量搜尋,令部分企業資料在毋須另行登入下外露。Claroc 檢查逾 40,000 個 grok.me 子域名,確認 37,571 個仍在線,其中 35,578 個(94.7%)可進入應用介面並存取資料。

Corporate data exposed through Grok Build-based applications, as obtained by security firm Claroc. An internal SpaceXSI document appears in the upper left. (Source: Ars Technica)
The rise of vibe coding, which enables users to generate code and applications using natural language, has made it easier for non-developers to build business software. However, weak default security designs in AI coding platforms, combined with inadequate corporate data management practices, have led to repeated incidents in which internal information is exposed on the internet.
Recently, data belonging to thousands of users was exposed through Grok Build, a vibe coding agent operated by SpaceXSI, the new name of xAI following its acquisition by SpaceX. The exposed data included proprietary business information and sensitive corporate secrets. According to technology publication Ars Technica, internal data belonging to SpaceXSI itself was also exposed.
Research by security firm Claroc, which first discovered the vulnerability, found that the exposed corporate data included internal business plans, advertising proposals and medical examination results. Some sensitive information originated from 24 applications developed by SpaceXSI itself. Documents found in these applications included a confidential onboarding plan for new enterprise sales employees, a sales territory map covering 594 customer companies, internal system addresses for sales tools and advertising proposals.
Claroc also discovered sensitive data belonging to other Grok Build users, including employees of Fortune 500 companies and defense contractors. A Peruvian construction company had an occupational health dashboard exposing medical examination records for 3,784 workers. At Tesla, a service center performance report containing the real names, productivity figures and bonus payments of 22 mechanics was accessible.
Other exposed information reportedly included personnel rosters for a U.S. Navy ship engineering contractor, a confidential proposal containing a hardware bill of materials (BOM) from a U.S. IT services company, and explosives inventory dashboards and handling procedures at the U.S. facility of a Norwegian-Finnish aerospace and defense group. The companies involved were not identified.
Claroc notified SpaceX of the findings on Nov. 8 but received no response. According to Ars Technica, SpaceXSI removed links to 24 applications created by its own employees and X employees within hours of receiving an email inquiry from the publication. However, Claroc said no separate access restrictions had been implemented for applications created by other users. At the time of Ars Technica’s report, all 24 applications displayed an “App not found” message.
Shared Link Vulnerability Exposes Apps Through Passive DNS
The unintentionally exposed data became accessible through links generated by Grok Build users to share their applications with others. Grok Build’s link-sharing feature is designed to simplify application development and deployment. Users can create internal testing applications within minutes using voice or text prompts and quickly distribute them to authorized users through generated links.
Grok Build generates randomized URLs intended to restrict access to stored data to people who know the exact address. Random word combinations make the URLs difficult for unauthorized parties to guess. This approach is also known as capability-based access control, in which possession of the URL itself grants access to the associated resource.
Each application receives a subdomain consisting of random words, such as “failing-chassis-traders.” Users can modify the subdomain, while the address ends in “grok.me.” Although this makes addresses difficult to guess, the links lose their secrecy if their hostnames are collected by passive DNS services.
The problem is that the randomized string appears in the hostname rather than the URL path. Whenever a user accesses an application, a DNS lookup occurs, potentially allowing the hostname to be collected by passive DNS services. Searching passive DNS databases for grok.me subdomains can therefore reveal large numbers of individual application addresses.
Claroc examined more than 40,000 grok.me subdomains discovered through passive DNS searches and confirmed that 37,571 were online at the time. Of those, 35,578, or 94.7%, allowed access to application interfaces and exposed data without requiring a separate login.
Claroc also reported that a week of testing with a paid account showed the default sharing permission for published applications was “Anyone with the link.” Unless users changed the visibility setting to “invited users” or “only you” during publication, anyone who knew the link could access the application simply by clicking it. Because the URL structure allowed hostnames to be collected through passive DNS, outsiders could discover these applications in large numbers.
According to Claroc, other vibe coding platforms that assign separate hostnames to individual applications, including Lovable, Bolt and Replit, may also allow application addresses to be discovered through passive DNS. However, whether data is actually exposed depends on each application’s visibility settings and implementation of authentication and access controls.
Corporate Vibe Coding Apps Fall Outside Security Governance
Data breaches caused by a combination of design weaknesses in vibe coding platforms and user negligence have become increasingly common.
According to a June report by the Cloud Security Alliance (CSA), The Vibe Coding Governance Gap, social network Moltbook exposed 1.5 million authentication tokens and 35,000 email addresses after embedding a Supabase API key in client-side JavaScript and disabling database row-level security (RLS). Consumer application Cal AI also exposed approximately 3.2 million user records, totaling 14.59GB, because of an unauthenticated Google Firebase backend.
In early May, Israeli security company Red Access examined approximately 380,000 publicly accessible assets associated with vibe coding platforms Lovable, Base44, Replit and Netlify. It found that about 5,000 exposed sensitive corporate and personal information, including medical records, financial data, corporate strategy documents and customer information.
Although it is not an AI coding agent, Meta Muse experienced a similar problem to Grok Build, with approximately 6,800 shared links exposed through passive DNS records. Claroc also discovered this issue, prompting Meta Muse to suspend its conversation-sharing feature on Oct. 20. Because conversation titles appeared in the hostnames of shared links, the links could be reconstructed from DNS records alone. Conversations were accessible without logging in, and some contained sensitive personal and business information.
AI coding platforms allow nonprofessional developers to create applications within minutes and share them through a single link, but security safeguards have taken a back seat, as demonstrated by public-by-default settings, weak access controls and URL designs that make addresses easy to discover and collect. Data breaches recur when users add employee lists, customer information and internal documents without security reviews or checking sharing permissions.
Vibe coding platforms should make private access the default and securely configure authentication and access permissions for individual applications. URL structures should also be improved to prevent easy discovery by outsiders, while authentication mechanisms must not rely solely on obscuring URLs.
Companies must also identify applications created by employees using AI and require those handling customer information and internal documents to undergo the same security reviews as conventional software, including authentication, access controls and data protection. Even in the era of vibe coding, where speed is a competitive advantage, security must be a fundamental requirement considered from the development stage rather than an optional feature added after deployment.
來源:The Elec · thelec.net