AI Lets Anyone Build Software, Leaving Security Behind
Claroc Research found 35,578 Grok Build apps accessible without login, exposing a structural gap as AI tools let non-developers ship unguarded software.

Artificial intelligence coding tools have made it routine for employees outside engineering teams to describe an application in a chat window and publish it to the web within minutes. Yet a November 10 investigation by Claroc Research shows what happens when software creation scales faster than access control: out of 37,571 live applications built with Grok Build under the grok.me domain, 35,578—or 94.7%—opened to any visitor with no sign-in required.
The findings do not point to an external breach of servers operated by SpaceXSI, the company formerly known as xAI. Instead, the exposure stems from how chat-based builders handle publishing defaults, subdomain DNS visibility, and generated authorization logic. As software generation shifts into the hands of sales representatives, hospital coordinators, and operations managers, platforms are letting non-developers ship functional web tools without enforcing basic security boundaries.
Inside the Exposed Corporate and Medical Apps
Most of the 35,578 unauthenticated pages were harmless experiments, calculators, landing pages, and small games, while only 1,993 live apps (5.3%) returned an HTTP 401 authentication error. According to AI Chat Daily, Claroc first ran an automated model over the saved landing pages that flagged 3,477 builds—roughly 9.8% of the open total—as potentially containing private material, before narrowing the estimate through a stricter second pass and a random 100-app audit to roughly 650 to 800 apps (about 2%) holding sensitive business or personal data, with 150 exposures confirmed by hand.
Even a narrow percentage across tens of thousands of deployments produced severe leaks spanning 30 identifiable external organizations, 26 of which employ more than 1,000 staff and six of which sit on the Fortune 500 list. Hand-confirmed examples included a hospital group app with a searchable directory of more than 3,500 employees, an intensive care patient briefing containing laboratory results, imaging, and ventilator status, and a contractor timesheet whose browser-loaded script stored 18 accounts with plain-text passwords, including administrators. Researchers also documented a fire department's live shift schedule with badge numbers, a cruise ship galley inspection checklist listing crew names and employee IDs, and a trucking firm's recruiting roster containing legal names and personal emails for over 1,000 driver candidates.
Twenty-four of the unauthenticated applications were created by employees of SpaceXSI and X themselves. Those internal builds included a confidential enterprise sales onboarding plan listing 19 internal chat channels, a sales territory map tracking 594 strategic accounts and their pipeline, eight X advertising proposals—among them one marked confidential for a U.S. federal agency—and four live advertiser boards, one monitoring progress toward a $1.6 million sales target.
How Unshared Links Surface Through Passive DNS
A central misconception around chat-generated links is that an obscure URL behaves like a password. Every application published through Grok Build receives its own subdomain under grok.me, often composed of four random words, and the moment an authorized recipient opens that address, the browser performs a DNS lookup that commercial passive DNS services record and index. Claroc extracted 43,601 grok.me hostnames through passive DNS queries between October 31 and November 8, 2026, testing 43,551 distinct addresses without ever receiving a link from an owner.
Claroc calls this pattern GhostShares, and it extends beyond Grok. On October 20, 2026, The Verge reported that Claroc had located more than 6,800 shared conversations from Meta's Muse AI assistant in public DNS records—exposing small-business loan applications, marketing revenue reviews, and interview preparation notes—before Meta spokesperson Tracy Clayton said the company "paused this functionality while we review this report." Earlier, in August 2025, more than 370,000 shared Grok chat links under grok.com/share/ had also appeared in Google search results because search crawlers could index those pages.
Architecture choices at the URL and default-permission layers determine how widely that exposure spreads. Claroc noted that published apps from ChatGPT also use dedicated hostnames (<app-name>.<username>.chatgpt.site) that enter passive DNS indexes once opened, as do apps deployed on Lovable, Replit, v0, Bolt, Base44, and Google AI Studio; however, when Claroc clicked Share on a ChatGPT app, access defaulted to "Only me" until changed manually, whereas shared conversations on Claude, ChatGPT, and Gemini place identifiers in the URL path so DNS resolvers never log them.
The Dispute Over Publishing Defaults
SpaceXSI pushed back on the framing that its platform automatically opens user projects to the public. Responding to media questions on November 10, the company stated that Grok apps remain private until a creator completes multiple publishing confirmations—clicking Publish, choosing a subdomain, selecting between "Only you" and "Anyone with the link," and accepting a legal disclaimer—and said employee accounts default to "Only you."
Claroc did not dispute that creators must click Publish, yet its testing pointed to a different design choice in that flow. On the paid account researchers used, the access selector already had "Anyone with the link" preselected after clicking Publish, requiring the user to opt out rather than opt in. As Asaf Aprozper of Claroc Research wrote, "the default answer to who can see it is everyone"—and despite SpaceXSI's stated internal setting for staff, 24 employee-built apps still ended up publicly readable until the company removed them following Claroc's November 8 and 9 notifications.
When Every Employee Becomes an Unaudited Developer
Default sharing settings are only half the problem when non-technical staff use prompts to assemble operational software. Tools such as Grok Build, Lovable, Replit, and Google AI Studio bypass traditional IT procurement and security reviews, allowing a lunch poll and a customer database to go live through the exact same interface while raising broader governance questions around unguarded AI tools.
Asking an AI model for login protection does not guarantee sound authorization code either. To verify whether open apps could also be modified without touching anyone else's live data, Claroc built two test applications with fabricated records: a public contact list with no sign-in where an incognito visitor added rows seen by subsequent visitors, and a customer tracker requested for a four-person sales team. For the sales app, Grok generated "Continue with Google" and "Continue with X" buttons, but the underlying code added any authenticated user to the team before checking membership, letting an uninvited Google account read, edit, and delete every customer entry.
Claroc confirmed that switching its first test app manually to "Only you" immediately returned HTTP 401 errors to outsiders, and unpublishing the second took it offline within seconds.
Direct Notifications and Unanswered Design Requests
Claroc Research submitted four architectural recommendations to SpaceXSI on November 8, urging the company to serve shared apps on randomized paths under a single host, introduce organization-wide domain restrictions, default new builds to "Only you," and trigger publish-time warnings when an app pairs a database with no authentication.
SpaceXSI removed the 24 staff-built applications after Claroc's disclosures, while external user apps identified in the scan remained online at publication and the company had not publicly committed to altering its consumer publishing flow. Because SpaceXSI did not confirm it would notify affected third-party builders, Claroc began contacting exposed organizations directly—reaching three before publishing its report—leaving corporate security teams auditing passive DNS logs for internal tools built outside corporate IT oversight.
Written by Adriano Nascentes
Founder & Lead AI Systems Analyst · Verified Staff
Founder and editor-in-chief of Unknown Observer. Technology researcher and software engineer investigating foundation models, autonomous agent runtimes, cognitive architectures, and frontier AI safety.
Was this article helpful?
Your feedback helps our editorial team refine reporting depth and accuracy.



