Grok Build flaw exposes patient records and internal SpaceXSI files

Akhil Ponnada
Akhil Ponnada is the founder and editor of FrontWord. He covers artificial intelligence, chips and infrastructure, consumer technology, cybersecurity and the business of tech.
7 Min Read
Image: Engadget

A Grok Build flaw has left more than 35,000 custom web applications open to the public without a login, exposing patient records, plaintext passwords and internal business documents from SpaceXSI and X. Cybersecurity firm Claroc reported the findings after inspecting tens of thousands of active apps hosted on grok.me.

Claroc analysed 37,571 applications that were still online and found that 35,578 required no authentication, a rate of 94.7 percent, according to the report. As The Decoder reported, the researchers estimate that between 650 and 800 of the accessible applications contained private personal data or confidential company records.

The security firm described the open web pages as GhostShares, because third parties could locate and open them without ever receiving a shared link. As Engadget reported, Claroc researchers wrote in a blog post: “The company that makes Grok had its own internal material open to anyone holding the link.”

SpaceXSI, formerly known as xAI, offers Grok Build as a chat assistant that writes and deploys software from plain text prompts. Creators make an app live by clicking a publish button, which generates a distinct web address hosted on the grok.me domain.

Claroc found that the publish dialogue on paid accounts preselected the option “Anyone with the link”. The service generated addresses using four randomly generated words with up to 38 bits of entropy. However, Grok Build placed those words as subdomains to the left of the domain name rather than as directory paths to the right.

Because each project ran on its own subdomain, the addresses registered in commercial passive DNS databases after their initial visit. Security researchers and potential attackers could then query those databases by pattern to compile lists of active apps without having to guess obscure URLs.

To fix the exposure, Claroc recommended that SpaceXSI host applications on hard-to-guess paths on a shared domain, matching the method used by Google’s Gemini. Placing the random string to the right of the domain prevents the unique addresses from leaking into passive DNS records.

What the exposed apps contained

The majority of the open projects were simple tools such as calculators, landing pages and small games. Even so, Claroc found that roughly one in ten contained sensitive business materials or personal records, and classified about two percent as severe security hazards.

Confirmed leaks included an intensive care unit briefing containing a patient’s laboratory values, an internal roster naming more than 3,500 staff at a hospital group, and an operational shift schedule for a fire department. In another case, the source code of a timesheet app revealed 18 user accounts with plaintext passwords.

The exposure also extended directly to SpaceXSI and social platform X, where Claroc identified 24 public applications created by company staff. These included a confidential sales onboarding guide containing internal network addresses and 19 private chat channels, alongside a strategic territory map tracking 594 corporate accounts.

Staff apps also revealed eight advertising proposals for X, including one marked confidential that targeted a US federal agency. Researchers found live advertiser account dashboards, campaign spending summaries and two copies of an internal brand safety playbook. Only one of the examined employee apps required a password.

Beyond SpaceXSI and X, the researchers identified 30 affected corporate organisations. Six of those businesses are listed on the Fortune 500, and 26 have workforces exceeding 1,000 employees. Claroc has withheld the names of the organisations while alerting administrators.

Flaws in Grok-written code

In tests conducted on its own accounts, Claroc discovered that Grok Build generated software with broken access controls. A contact list application built by the model created a backend database that lacked login controls entirely, allowing anonymous visitors to insert records.

When researchers instructed Grok to build a customer directory with user logins, the assistant added Google and X sign-in options. The generated code automatically enrolled every authenticated visitor into the team before verifying permissions, allowing any external Google account to view, alter and delete stored entries.

Grok Build provides an “Only you” setting that restricts access to the creator, but users must choose it manually. Claroc observed that OpenAI’s ChatGPT defaults to an owner-only setting when sharing projects. Claroc urged SpaceXSI to make restricted access the default and display explicit permission warnings before publishing.

SpaceXSI response and past privacy issues

SpaceXSI defended the platform by stating that applications remain private until a user explicitly publishes them through a workflow that includes a legal notice. The company added that accounts assigned to its own employees default to the “Only you” setting.

Claroc argued that multiple confirmation screens provide little protection when “Anyone with the link” remains selected at the final step. The firm offered to share its full list of exposed addresses confidentially, but SpaceXSI declined to commit to contacting affected users.

SpaceXSI removed the 24 internal employee applications only after Claroc tagged Elon Musk and a company executive on X. The company has not taken down the thousands of apps left open by external users and organisations, prompting Claroc to contact the owners directly. Claroc noted that it sells security products for AI-built software.

The discovery adds to earlier data handling issues at SpaceXSI. In August 2025, Forbes reported that search engines had indexed more than 370,000 shared Grok conversations. In July 2026, the Grok Build terminal utility uploaded local files, including SSH keys and password databases, to company servers without explicit confirmation.

Regulators are also scrutinising the company. Ireland’s Data Protection Commission is investigating whether X neglected data protection by design following the generation of sexualised deepfakes by Grok.

Sharing mechanisms across the AI industry have suffered comparable leaks. In October, Claroc identified 6,800 shared conversations from Meta’s Muse agent using the same passive DNS discovery technique, while shared chats and artifacts from Anthropic’s Claude appeared in public search engine results earlier in the year.

Follow:
Akhil Ponnada is the founder and editor of FrontWord. He covers artificial intelligence, chips and infrastructure, consumer technology, cybersecurity and the business of tech.