top of page
40051abd5a76713af8f015988fc6780e-blue-phone-icon-with-a-wave-on-it.webp
Logo-Zalo-Arc.webp
Facebook_Messenger_logo_2020.svg.webp
whatsapp

Over 35,000 Grok apps publicly accessible: What businesses need to check

5 days ago
4 min read

An employee builds a customer dashboard with AI, adds real business data, and shares the link with colleagues. The application works. But has anyone checked whether an outsider can open it?

Research covered by The Hacker News on November 11, 2026, puts that question in focus. Thousands of applications built with Grok were accessible without authentication, with some exposing internal business information and personal data.

For businesses using AI app builders, the practical response is to establish which applications exist, what information they contain, and whether their access controls work.

What did researchers find?

In its GhostShares 2 report, security company Claroc said it identified 37,571 Grok applications that remained online during its investigation. Of those, 35,578 - approximately 94.7% of that sample - could be opened without signing in. This percentage describes the researchers’ sample, rather than all Grok applications.

grok-apps-public-access-business-data-security
Over 35,000 Grok apps publicly accessible

The Hacker News reported that most were games, landing pages, and calculators. However, some exposed confidential business documents, employee information, patient details, or plaintext passwords.

Public accessibility does not establish that every application was compromised or contained sensitive information. Exposure and confirmed exploitation require separate evidence.

Claroc also reported that publishing required user action. In the account it tested, “Anyone with the link” was preselected during publishing. The provider said employee accounts defaulted to private access. These findings should therefore be understood in the context of the tested account and the provider’s response.

Why a shared link is not an access control?

Grok’s product announcement describes applications published to individual addresses under grok.me, with options controlling who can open them.

Claroc explained that application hostnames could be discovered through passive DNS records. It also demonstrated an authorization problem in an application it created: an uninvited account could sign in and interact with customer records. The researchers conducted modification and deletion tests only on their own applications.

The broader security principle is straightforward: authentication establishes identity; authorization determines what that identity can access.

A Google sign-in button may identify a visitor without establishing that the visitor belongs to the intended team. OWASP recommends denying access by default and validating permissions on every request.

Businesses evaluating vibe coding security risks should therefore require evidence that unauthorized access is blocked, alongside evidence that the application functions correctly.

Which business applications deserve attention first?

Start with applications handling real customer, employee, financial, or operational information.

The following are hypothetical review scenarios, rather than confirmed incidents from the Grok investigation.

Application

Security question

Customer dashboard

Can an outside account view or export customer records?

Employee scheduling tool

Can staff access information beyond their assigned role?

Financial reporting application

Who can read reports, alter figures, or download files?

Supplier portal

Are suppliers restricted to their own organization’s records?

Application connected to another service

Where are credentials stored, and who can revoke them?

A useful review begins with the data and business process. A simple interface may still provide access to commercially sensitive records.

5 checks before an AI-built app goes live

1. Assign an accountable owner

Record who owns the application, which department uses it, and who controls its administrative account.

The inventory should include its purpose, hosting platform, data sources, intended users, and retirement process. Someone must be able to restrict or remove the application when its creator changes roles or leaves.

Applications introduced outside established approval processes should be included in the organization’s Shadow AI review.

2. Test permissions across user roles

Within an authorized testing scope, check access using:

  • A visitor who has not signed in.

  • An external account without an invitation.

  • A legitimate user with limited permissions.

  • A user whose access has been revoked.

Document whether each account can read, change, delete, or export information.

Permission checks must apply to server requests as well as visible interface controls. Hiding a button does not establish that the underlying action is restricted. This follows OWASP’s guidance to validate authorization on every request.

3. Approve the use of real data

Use synthetic or appropriately de-identified information during early development.

Before adding production data, establish who can approve its use and what conditions the application must meet. Review those conditions again when the application gains a new data source, integration, or audience.

For example, approval for a prototype using fictional customer records should not automatically cover a later connection to the live CRM.

4. Protect credentials and service keys

Keep passwords and service credentials out of code delivered to the browser.

If a credential may have been exposed, revoke or replace it at the issuing system. Removing it from the current application does not invalidate copies that may already exist.

OWASP’s secrets management guidance supports revoking potentially compromised secrets and managing their replacement throughout their lifecycle.

5. Set security acceptance criteria

Define the security checks required before an application handles important business data.

The assessment should reflect its functions, users, integrations, and potential impact. For higher-risk applications, web application penetration testing can provide an independent assessment within an agreed scope.

Reviewers need access to the relevant user roles and business workflows to test whether the application enforces its intended boundaries.

What if an internal application is already public?

Restrict access or suspend publication while preserving available logs and configuration evidence.

Then establish what information was accessible, which credentials may need replacement, and which connected systems require review. Verify the corrected controls before restoring access.

Incident reporting should distinguish confirmed findings from uncertainty. If historical logs are insufficient to establish who accessed the application, record that limitation.

Privacy and legal teams should assess any notification obligations against the affected data, contracts, and applicable jurisdictions.

Make application review part of AI governance

An initial review should lead to an ongoing process.

Register applications when they are created, approve their use of sensitive information, and reassess them when their functionality or audience changes. These responsibilities can sit within the organization’s AI governance process.

The immediate question for business leaders is practical: can your organization identify its AI-built applications and demonstrate who can access their data?

Start with applications already handling real information. Confirm their owners, review sharing settings, and test the permissions that protect their most important workflows.

Need to assess an AI-built business application? Contact IPSIP Vietnam to discuss a security review focused on authorization, sensitive data flows, and credential management.

-----

References

follow ipsip vietnam.png
bottom of page