Skip to article
Datapath abstract technology background
Back to News
CYBERSECURITY • Published November 15, 2026 • 7 min read • Source: Claroc Research, November 10, 2026

Grok Build Exposure Shows Why AI Needs Secure Guardrails

Grok Build's reported exposure shows why AI apps need private sharing, tested permissions, scoped credentials, human approval, and ongoing monitoring.

Dan J Sturdivant, Vice President at Datapath

By

Vice President

cybersecuritydata securitymanaged IT

Key takeaways

  • Claroc reported open Grok apps containing confidential material, while noting that most reachable apps were ordinary public apps.
  • Secure AI deployment requires tested access controls, limited data and credentials, and enforced approval for consequential actions.
  • Treat publication as a release decision with an accountable owner, monitoring, and a tested shutdown procedure.

Original source

Claroc Research, November 10, 2026

What does the Grok Build exposure mean for secure AI deployment?

Secure AI deployment requires controls over what an app can access, who can use it, and what happens when someone publishes it. Businesses should require approved data, authenticated sharing, server-side permission checks, and a tested shutdown procedure before an AI-built prototype becomes a tool that staff or customers depend on.

The practical lesson for leadership is to evaluate the deployed application, not just the reputation of the company supplying the model. A convincing demonstration shows that software works. A security review establishes who it works for, what it exposes, and how its access can be withdrawn.

What did researchers actually find?

Claroc’s November 10 research reported that 35,578 of 37,571 reachable Grok apps in its sample opened without sign-in. Most were ordinary public apps; that count does not mean 35,578 sensitive breaches. The researchers identified 24 open apps attributed to SpaceXSI and X staff, including confidential business material, and said those apps were subsequently removed.1

Claroc found app addresses through passive DNS records. In its tested account, publishing proposed “Anyone with the link.” It also demonstrated unauthorized record changes after an uninvited account signed into an app the researchers created themselves. It did not test writes against other people’s apps.1

According to the vendor response reproduced by Claroc, apps remain private until users publish them, and employee accounts default to “Only you.” The dispute concerns protection after publication. The disclosure demonstrates exposure and a test-app authorization failure; it does not establish criminal exploitation of the exposed apps.1

The reported exposure of the provider’s own business material is a warning against assuming AI expertise guarantees secure everyday deployment.

SpaceXSI’s own product announcement describes private, link-based, and public sharing, alongside business-data connectors and storage for third-party API keys outside application code. Those capabilities make configuration and verification part of the deployment decision.2

Which guardrails should be in place before an AI-built app goes live?

The minimum release standard should cover identity, data, permissions, secrets, and operational ownership. The controls must operate in the application and its supporting systems. A prompt asking the AI to make something secure can help describe the requirement, but the release decision needs evidence that the requirement is enforced.

1. Make private access the starting point

For an internal tool, approve named users or managed groups before sharing the application. Require the organization’s identity provider where supported, with the appropriate multifactor-authentication policy. Public access should be a deliberate exception for content that has been approved for public release.

Treat a preview link as a distribution mechanism, not an access-control policy. Ask the application owner to explain what happens when an employee forwards the link, leaves the company, or loses access to the underlying business system. Those scenarios belong in the acceptance test.

2. Test authorization after sign-in

A successful login establishes identity. The application still needs to decide which records that person may read, change, export, or delete. OWASP recommends configuring applications to “deny access by default” and checking permissions on every request.3

Test with an invited user, an uninvited user, a signed-out browser, and a user whose access has been revoked. Include the underlying data endpoints, not only the screen containing the sign-in button. For systems serving multiple customers, verify separation between customers using synthetic records and authorized test accounts.

Keep the results with the release record. A screenshot of a login page cannot establish that the data behind it is protected.

3. Limit the information and credentials the app receives

Build the first version with synthetic data. Before introducing real records, identify the fields the workflow actually needs and remove the rest. A scheduling tool may need availability without needing a complete personnel file. A sales dashboard may need totals without exposing customer contact details to every viewer.

Store secrets in an approved server-side secrets service. Avoid embedding passwords, privileged API keys, or database credentials in browser-delivered code. Use credentials scoped to the application’s specific job, and separate production access from development access.

OWASP’s secrets guidance covers centralized management, least privilege, auditing, expiration, and revocation. If a credential has been exposed, removing it from the page is only part of the response: revoke or rotate it at the issuing service and verify that the old value no longer grants access.4

4. Separate building from publishing and consequential actions

Give employees a quick way to experiment without giving every prototype immediate production access. A practical progression is a synthetic-data prototype, a restricted pilot, and a reviewed production release. Each stage should have an owner and a clear decision about the next permission being added.

For agents that can send messages, alter business records, deploy software, or delete data, require an appropriate approval step before the action executes. The enforcement belongs in the execution layer, outside the model’s own reasoning. OWASP’s agent-security guidance recommends scoped tools, explicit approval for high-impact actions, action previews, and audit trails.5

For example, an assistant can prepare a customer update while a staff member approves the recipients and final text. The assistant does not need unrestricted authority to email the entire CRM because it can draft a useful paragraph.

5. Recheck the app when its purpose changes

A harmless calculator can become a business application after someone adds a spreadsheet, a database, or a connector. Require a fresh review when the data classification, audience, permissions, or integrations change. Keep a record of the approved configuration so the next reviewer has something concrete to compare.

Monitor meaningful events: public-sharing changes, new integrations, failed access checks, unusual exports, and changes to privileged roles. Avoid collecting passwords or full sensitive records in logs. Test that the team can disable sharing, revoke a connector, and restore the last approved version without depending on the original builder being available. OWASP’s logging guidance provides a useful reference for security events and information that should be excluded from logs.6

What should teams do if they already use AI app builders?

Start with an inventory and review the highest-risk applications first. Ask departments what they have built for real work, including informal dashboards, rosters, trackers, and customer tools. Give employees a straightforward way to report a prototype so IT can help secure it before it becomes an operational dependency.

For each application, record its owner, audience, data categories, publishing setting, connected systems, and shutdown procedure. Prioritize anything containing health information, employee records, financial information, customer data, or credentials.

If a review finds unintended public access:

  1. Restrict access or unpublish the affected application while preserving the evidence needed for investigation.
  2. Identify which records and credentials were accessible and during what period.
  3. Revoke exposed credentials and review activity in the connected services.
  4. Involve the organization’s security, privacy, and legal owners to assess notification duties.
  5. Reopen the workflow only after access tests pass and a responsible owner approves restoration.

Record uncertainty honestly. Missing logs can leave the team unable to determine who accessed information. That is a limitation to address in the response, rather than evidence that nobody did.

How can leaders make these controls practical?

Put the rollout decision in one short record: the business benefit, approved users and data, technical controls, test results, accountable owner, and conditions that would trigger a pause. This gives management a useful decision without making every experiment a large procurement project.

The voluntary NIST AI Risk Management Framework provides a structure for evaluating and managing AI risks throughout use. It can support this process, but adopting the framework does not certify a tool or establish regulatory compliance.7

Datapath’s secure AI adoption roadmap provides a staged approach, while our third-party access review guide helps teams examine the permissions behind a vendor relationship.

Datapath supports this work through AI governance consulting and co-managed IT services, connecting policy, vendor review, identity controls, and operational responsibility.

Preparing an AI rollout?

Talk with Datapath about the data, access tests, and approval steps your first production workflow needs.

Discuss your AI rollout

Sources

Footnotes

  1. Claroc Research, Asaf Aprozper, “GhostShares 2: Grok Apps Exposed xAI’s Business Secrets and People’s Private Data,” November 10, 2026. Research findings and reproduced vendor response; rechecked November 14. ↩ ↩2 ↩3

  2. SpaceXSI, “Grok Build on web and mobile,” August 19, 2026. ↩

  3. OWASP, Authorization Cheat Sheet. ↩

  4. OWASP, Secrets Management Cheat Sheet. ↩

  5. OWASP, AI Agent Security Cheat Sheet. ↩

  6. OWASP, Logging Cheat Sheet. ↩

  7. NIST, AI Risk Management Framework. Framework released January 26, 2023; current overview rechecked November 14, 2026. ↩

Disclaimer: This news summary is intended for informational and marketing purposes only, and nothing presented here is contractually binding or necessarily the final opinion of the authors.

Need to turn industry change into an IT plan?

Datapath can help translate security, compliance, and infrastructure signals into practical next steps for your organization.

Book an IT Consultation