What happened?
On November 10, 2026, cybersecurity research firm Claroc published a report titled GhostShares 2 describing security weaknesses in applications created using Grok Build.
Grok Build is an AI-powered application development environment from SpaceXSI.
The platform allows users to describe an application in natural language and have Grok generate the underlying code.
Claroc discovered that a substantial number of published applications could be opened without requiring a login.
Its investigation also identified confidential material belonging to SpaceXSI itself, along with sensitive information associated with other organizations.
The incident demonstrates how quickly AI-generated applications can create security risks when access controls and sharing configurations are not handled correctly.
1. More than 35,000 applications accessible without authentication
Claroc investigated application addresses associated with the grok.me domain.
Its report provides the following figures:
| Research metric | Reported result |
|---|---|
| Application addresses tested | 43,551 |
| Applications still online | 37,571 |
| Applications accessible without login | 35,578 |
| Percentage accessible without login | 94.7% |
| SpaceXSI and X employee applications identified | 24 |
| Other organizations identified | 30 |
| Estimated applications containing private data | 650–800 |
These figures come from Claroc's own investigation.
Importantly, the 94.7% figure does not mean that every accessible application contained sensitive information.
Claroc estimated that approximately one in ten open applications contained sensitive, internal or personal information.
2. What information was exposed?
The investigation identified multiple examples of sensitive information.
Some of the exposed applications reportedly contained:
- Confidential internal onboarding documents.
- Enterprise sales territory information.
- Advertising proposals.
- Employee names and identification numbers.
- Medical examination records.
- Hospital staff directories.
- Employee schedules.
- Plaintext passwords stored in application source code.
One particularly significant discovery involved 24 applications created by SpaceXSI and X employees.
These included an internal onboarding plan and a sales territory map containing information about 594 strategic accounts.
Claroc also identified advertising proposals, including one marked confidential and prepared for a US federal agency.
The investigation found information associated with 30 other organizations, six of which were Fortune 500 companies.
The companies were not all publicly identified in the research report.
3. How did the security weakness occur?
The central issue involved the way Grok Build handled application sharing.
When an application is published, the platform creates a dedicated web address.
The intended functionality is to allow users to share their applications through these links.
However, Claroc found that the publishing interface in the paid account it tested had the option:
Anyone with the link
preselected.
The platform also provides more restrictive options, including:
- Only you.
- Invited users.
- Anyone with the link.
The researchers argued that the public-sharing default created a significant risk for users who might not understand the implications of publishing their applications.
The passive DNS problem
Another important technical detail involved passive DNS.
When a browser opens a domain, it performs a DNS lookup to discover the server associated with that address.
Passive DNS databases collect and record information about these lookups.
Grok Build generated application-specific subdomains.
Because the randomized identifiers appeared before the grok.me domain, researchers could search passive DNS records to identify application addresses.
This made the links more discoverable than users might have expected.
A supposedly difficult-to-guess URL should not be treated as a replacement for authentication or authorization.
Even if someone does not know the original sharing link, the hostname may become discoverable through external DNS records.
4. A second problem: AI-generated authentication code
Claroc also tested two applications it created itself.
In one experiment, a generated contact-list application contained a database without authentication.
In another, an application designed for a sales team included Google and X sign-in.
However, the generated application code added authenticated visitors to the team before checking whether they were actually authorized members.
This meant an outside Google account could access and modify records in that particular test application.
This is an important distinction.
The researchers demonstrated this behavior in their own test application. It should not be interpreted as proof that every Grok Build application contained the same authentication flaw.
5. How did SpaceXSI respond?
SpaceXSI provided a response disputing the interpretation of the default-sharing behavior.
The company said that applications remain private until creators choose to publish them, that publication requires multiple steps and that employee accounts default to the more restrictive setting.
Claroc argued that its tests showed the final sharing selection was still preconfigured to allow anyone with the link.
There is therefore a disagreement between the company's description of its safeguards and the researcher's observed publishing behavior.
According to Ars Technica, SpaceXSI began removing its own 24 exposed employee applications after the publication's inquiry.
The company had not removed the broader population of other users' accessible applications at the time described in the research coverage.
Claroc also began contacting affected organizations directly.
6. Why this matters for developers and SaaS companies
This incident is particularly relevant to the growing popularity of vibe coding.
AI application builders make it possible to generate functional applications without manually writing every component.
However, application generation is only one part of software engineering.
A production-ready application also requires:
- Authentication.
- Authorization.
- Database security.
- Role-based access control.
- Secure configuration.
- Input validation.
- Secret management.
- Logging and monitoring.
- Security testing.
An application can look professional and function correctly while still exposing its database or internal business information.
This is why AI-generated code must be reviewed before being deployed.
Recommended security practices
| Security measure | Recommended implementation |
|---|---|
| Authentication | Require authentication for private business applications. |
| Authorization | Check user permissions on the server, not just the frontend. |
| Database security | Enforce database-level access policies. |
| Sharing defaults | Make private access the default. |
| Secret management | Never store passwords or API keys in frontend code. |
| Testing | Test applications using unauthenticated accounts. |
| Deployment | Separate development, staging and production environments. |
| Monitoring | Log unusual access and administrative activity. |
For SaaS developers, the biggest lesson is that an AI-generated application should undergo the same security review as traditionally written software.
7. What to watch next
The next important developments include:
- Changes to Grok Build's default sharing configuration.
- Additional authentication and authorization controls.
- Notifications to affected organizations.
- Independent verification of the exposed applications.
- Security updates to AI-generated application workflows.
- Broader security standards for AI application builders.
This incident could also influence how competing AI development platforms approach application publishing and default privacy settings.
Why this is trending
Ars Technica published detailed coverage of the investigation on November 10, including exposed internal documents and the passive DNS problem.
The Decoder published independent reporting on November 11 covering the 35,578 accessible applications, patient data and plaintext passwords.
Claroc Research published the original GhostShares 2 investigation on November 10.
Ars reported that Claroc contacted SpaceXSI and publicly tagged company representatives on X. The original posts were not independently accessible during the research run.
No sufficiently verified same-day Hacker News or Reddit discussion was identified.
The story has substantial cybersecurity relevance even without claiming that it went viral across every platform.
Editorial verification
Editorial verification note: Claroc reports both approximately one in ten open apps with sensitive information and an estimated 650–800 apps containing private data. Those figures are not numerically equivalent, and the report does not clearly reconcile their scopes. The 94.7% figure concerns access without login, not sensitive-data prevalence. Independent news coverage describes Claroc’s findings; it does not constitute an independent replication of the scan.
Sources
- November 10, 2026 — Claroc Research: GhostShares 2 original security investigation.
- November 10, 2026 — Ars Technica: Grok Build coding agent leaks confidential information.
- November 11, 2026 — The Decoder: Grok-built apps exposed patient data and business information.
Further reading
Building a product in this space?
MORE FROM THE NEWSROOM
Related news
CybersecurityMalicious Optimizely Scripts Reached 178 Organizations, Exposing Website Visitors to Data Theft
A malicious JavaScript campaign distributed through Optimizely Web reached visitors across 178 organizations, according to security researcher Scott Helme. The payload included cryptocurrency address substitution, payment-card skimming, cookie theft and iOS exploit delivery mechanisms, highlighting the security risks associated with trusted third-party website scripts.
CybersecurityUber Driver Information Exposed After Fireworks AI Vendor Breach, State Filings Reveal
Uber has confirmed that personal information belonging to drivers was accessed through compromised credentials at a third-party technology provider identified as Fireworks AI. Public breach notices and independent reporting identify at least 7,870 affected individuals across Massachusetts, South Carolina and Rhode Island, with additional state notifications still being assessed.
CybersecuritySignDrop Windows Zero-Day Exposes Local SYSTEM Privilege-Escalation Path; Microsoft Investigates
Security researcher Abdelhamid Naceri has disclosed SignDrop, a Windows privilege-escalation vulnerability involving the Common Log File System and Microsoft Defender. A public proof of concept demonstrates a path from an existing local user account to SYSTEM privileges, although Microsoft says Defender detects and blocks the associated exploit behavior and that it has not observed exploitation in the wild.
