Back to list
Nov 12, 2026
3.0K
0
0
Other LLM

Grok Build Apps Exposed: 35,578 Open Links Found by Claroc

Claroc tried 43,551 grok.me app addresses; 35,578 opened with no sign-in, including 24 built by SpaceXSI and X staff.

#Grok#SpaceXSI#xAI#Grok Build#Claroc
Grok Build Apps Exposed: 35,578 Open Links Found by Claroc
AI Summary

Claroc tried 43,551 grok.me app addresses; 35,578 opened with no sign-in, including 24 built by SpaceXSI and X staff.

Introduction

On November 10, 2026, the security firm Claroc published research it calls GhostShares 2, showing that apps people built and published with Grok's Build mode were openly reachable on the internet. Each published app gets its own address under grok.me, and Claroc says it tried 43,551 of them. Of the 37,571 still online, 35,578 (94.7%) opened for anyone, with no sign-in. Twenty-four of the open apps were built by staff of SpaceXSI, the company Claroc describes as formerly xAI and Ars Technica describes as the merged SpaceX and xAI, and of X. Ars Technica covered the research the same day and rounded the count to "35,000." This article analyzes the exposure itself, not the Grok Build coding model reviewed earlier on Evermx.

How the Exposure Works

On Grok, Claroc explains, the link is the hostname: every published app lives at its own address of the form name.grok.me. When anyone opens one, their computer asks a DNS resolver where that name lives, and passive DNS services collect those lookups and let account holders search them by pattern. Claroc says about half the names it saw were four random words, which nobody would guess, "and nobody needs to, because the dataset lists them." Ars frames the design as a variant of capability-based access control, where possessing the URL is the authorization, and notes that Google-style links keep the random part to the right of the domain, where DNS never sees it.

Claroc compared other platforms. Shared chats on Claude, ChatGPT and Gemini keep the secret in the URL path, so DNS lookups never reveal it. Claroc says Lovable, Replit, Base44, Google AI Studio and Meta's Muse, among others, also give each app its own hostname, which is what makes a name findable. Claroc says an app it built in ChatGPT was set to Only me until it changed it. The study scanned only grok.me, so it does not measure exposure on other builders.

Measure (Claroc)Count
Names listed under grok.me by subdomain discovery43,601
App addresses tried43,551
Still online37,571
Opened for anyone with no sign-in35,578 (94.7% of online)
Returned HTTP 401, the response of an app set to Only you or Invited users1,993 (5.3%)
Open apps built by SpaceXSI and X staff24
Other identifiable organizations with internal data in open apps30

What Was Exposed

Most open apps, Claroc says, were games, landing pages and calculators. About 1 in 10 held sensitive material, internal business data or personal data, and by Claroc's estimate roughly 2% held material serious enough to rate high or critical. It reached that figure by having an AI model flag pages that looked private, confirming them by hand and checking the model against a random sample, which put the total at roughly 650 to 800 apps with private data.

The 24 SpaceXSI and X apps included a confidential onboarding plan for a new enterprise sales hire, a sales territory map of 594 accounts, eight advertising proposals and media plans (one to a U.S. federal agency, marked confidential), four live advertiser account boards and seven campaign reports for advertisers in several countries. Among confirmed examples from other users, Claroc lists a hospital group's staff roster of more than 3,500 employees, an intensive care briefing for one patient, a fire department's live shift board and a contractor timesheet whose code held 18 accounts with plain-text passwords. Claroc identified 30 further organizations, 26 with more than 1,000 employees, and named none. Ars adds that affected users included Fortune 500 companies and a defense contractor.

What Claroc Tested and What It Inferred

Claroc loaded the first page of each app once, signed in to nothing, and says it did not sign in to, write to or test anyone else's app. Whether a visitor can also change data was tested only on two apps Claroc built itself, with made-up data. In the first, a contact list saved "so every visitor sees them" got a database and no sign-in, and a row added from an incognito window was visible to the next visitor. In the second, a team customer list, Grok added Google and X sign-in on its own, but the generated code added whoever signed in to the team before checking membership. A Google account that had never been invited read the customers, changed one record and deleted another.

Claroc proved these write and sign-in problems only on its own two apps, so whether other open apps allow edits, not just reading, has not been tested. Switching the first app to Only you made it answer 401 to a stranger, and unpublishing the second took its address offline within seconds.

Disclosure Timeline and SpaceXSI's Response

Date (2026)Event as reported
November 8Claroc reports to SpaceXSI at its security.txt address, including two of SpaceXSI's own open apps
November 9Claroc reports the other 22 SpaceXSI and X staff apps
November 10Claroc says SpaceXSI answered a reporter's questions but had not replied to Claroc's report, and had taken down the 24 staff apps but not other users' apps; Claroc begins notifying affected organizations and has emailed three by publication
November 10Ars reports that SpaceXSI began removing the links within a few hours of Ars emailing questions, and that all 24 returned "app not found" when the story went live

Claroc reports SpaceXSI's position as follows: apps are private until their creator chooses to share one, publishing takes several steps, and employee accounts default to Only you. Claroc answers that in the paid account it used, the access panel offered Anyone with the link, Invited users and Only you, with Anyone with the link already selected, and that the 24 open staff apps were built by SpaceXSI's own people. Ars quotes a representative it identifies as from Cursor, which Ars says SpaceXSI recently acquired, saying Grok Build apps "are private by default and nothing goes public unless the creator chooses to share it." Ars says Claroc's tests contradict that statement, because Grok Build required users to change the preselected Anyone with the link option, and that the representative did not answer follow-up questions, including whether SpaceXSI planned to notify customers who exposed similar data. Claroc describes the default as "a design choice SpaceXSI defended rather than a bug it said it would fix."

Practical Guidance

Claroc's checklist for teams: find out which AI app builders people use, search passive DNS for those domains, set any Grok app holding internal material to Invited users or Only you, test sign-in with an outside account, check script files for names, emails and passwords, and unpublish apps nobody owns. Claroc also sent SpaceXSI four proposed fixes, including a hard-to-guess path under one host, an organization-only option, Only you as the default, and a publish-time warning when an app has a database but no sign-in. Because Claroc sells software for securing AI-built apps and invites readers to book an exposure check, its recommendations should be weighed with that interest in mind.

Pros and Cons of the Research

Strengths: countable figures with a stated method, withheld hostnames and own-app write tests, and remedies builders can apply at once.

Limitations: it covers only grok.me, sensitive-app totals are estimates from model-assisted flagging, and Claroc is a commercial vendor in this space.

Outlook

Neither source reports a change to the default sharing setting for ordinary users or a SpaceXSI notice to the owners of the remaining open apps. Claroc says SpaceXSI did not say it would reach those owners, which is why Claroc began contacting organizations itself. Ars notes that Meta's Muse recently suspended a link-sharing feature after similar findings, and Claroc says its previous GhostShares post covered Muse. In Claroc's words, "AI app builders turned publishing into a few clicks, and the default answer to who can see it is everyone."

Conclusion

The documented facts are narrower than the headline and still serious. Claroc counted tens of thousands of open apps, found internal SpaceXSI material, and showed on its own test apps that missing controls can allow edits. SpaceXSI's account of a deliberate publish flow stands alongside Claroc's finding that the access step preselected sharing with everyone. Teams that let staff build apps with Grok should audit their published apps and set access to Only you or Invited users.

Editor's Verdict

Grok Build Apps Exposed: 35,578 Open Links Found by Claroc is a credible, carefully scoped piece of security research and a warning that applies to anyone publishing apps from an AI builder.

The strongest case for paying attention: the study gives countable figures with a stated method rather than anecdotes. Reinforcing that, a write-up that withholds hostnames and app titles and limits write tests to the researchers' own apps reduces further harm — a disclosure standard worth noting. The broader signal worth registering is straightforward: the exposure came from hostname design, since each app's address sits in a DNS name that passive DNS services index, rather than from any break-in. On the other side of the ledger, one constraint is real rather than a marketing footnote: the research covers only grok.me, so it cannot say how Grok compares with other AI app builders. Layered on top of that, its totals of sensitive apps are estimates from model-assisted flagging and manual review, not an exhaustive audit — so the headline count of open apps is firmer than the count of sensitive ones.

For security teams and any organization whose staff build and publish apps with Grok or similar AI app builders, the practical step is to audit published apps now and set access to Only you or Invited users rather than waiting for a platform-side fix.

Advertisement

Pros

  • The study gives countable figures with a stated method rather than anecdotes
  • A write-up that withholds hostnames and app titles and limits write tests to the researchers' own apps reduces further harm
  • Its findings come with concrete remedies that builders can apply immediately, such as setting access to Only you
  • Ars Technica reports the 24 SpaceXSI app links were still exposing data on Monday evening and returned app not found by publication

Cons

  • The research covers only grok.me, so it cannot say how Grok compares with other AI app builders
  • Its totals of sensitive apps are estimates from model-assisted flagging and manual review, not an exhaustive audit
  • Claroc is a commercial vendor in this space and its post promotes an exposure check
  • The default-setting finding comes from Claroc's own paid account, and SpaceXSI describes the publish flow differently
Advertisement

Comments0

Key Features

1. Hostname-as-secret design: each published app lives at its own address under grok.me, so passive DNS records make the names searchable 2. Scale: of 37,571 online apps Claroc tried, 35,578 (94.7%) opened with no sign-in 3. Contents: 24 open apps from SpaceXSI and X staff plus internal data from 30 other identifiable organizations 4. Own-app tests: on two apps Claroc built itself, anonymous visitors could write data, and a non-invited Google account could read, change and delete records 5. Disclosure: reported November 8, 24 staff apps taken down, other users' apps left open per Claroc

Key Insights

  • The exposure came from hostname design, since each app's address sits in a DNS name that passive DNS services index, rather than from any break-in.
  • In Claroc's data, 94.7% of online apps opened without sign-in, and the access panel in its own paid account had Anyone with the link already selected.
  • SpaceXSI's own staff built 24 of the open apps, so the exposure reached the company's internal sales and advertising material.
  • Write and sign-in weaknesses were demonstrated only on Claroc's two test apps, so edit access on other open apps remains untested.
  • SpaceXSI says apps are private until a creator shares them and that publishing takes several steps, while Claroc says the access step preselects sharing with everyone.
  • Claroc says SpaceXSI took down the 24 staff apps but, as of Claroc's post, not the apps of other users, and did not say it would reach their owners.
  • Claroc's comparison shows chat share links keep their secret in the URL path, while per-app hostnames expose names to DNS on several builders.
  • Claroc sells software for securing AI-built apps, so its remedies and its estimates deserve independent corroboration.

Was this review helpful?

Share

Twitter/X
Advertisement