Brief IA
›

Grok: 94.7% of Applications Open, SpaceXSI Removes 24 Links

Models & LLM·Tom Levy·

Grok: 94.7% of Applications Open, SpaceXSI Removes 24 Links

Grok: 94.7% of Applications Open, SpaceXSI Removes 24 Links
Key Takeaways
1Claroc identified 35,578 Grok applications accessible without login out of 37,571 verified
2Internal documents from SpaceXSI, personal data, and plaintext passwords have been exposed
3SpaceXSI removed 24 employee applications but not those of other users
Why it matters : Default configurations and ease of discovery massively expose sensitive data through AI-generated applications.

Dozens of thousands of Grok applications are accessible without a connection, containing internal documents and personal data, according to Claroc. SpaceXSI claims that going online requires several validations and says it has removed 24 applications from employees, while other cases remain open.

Limited Withdrawals and Privacy Precedents

SpaceXSI has removed 24 applications uploaded by employees, but has not deleted those from other users and organizations that remain accessible. This marks at least the third privacy incident involving Grok. In August, over 370,000 Grok conversations were identified as indexed by Google. In July, the Grok Build terminal tool had uploaded user files, including SSH keys and password databases, without prior solicitation. The Irish Data Protection Authority is also investigating X regarding sexualized deepfakes generated by Grok. Similar issues exist elsewhere: Claroc reports having spotted around 6,800 shared conversations from Meta's Muse agent in October, and shares from Anthropic's Claude were referenced in July. Claroc also notes that it markets security solutions for AI-built applications.

Twenty-Four SpaceXSI Employee Applications Published Online

Twenty-four accessible applications came from employees of SpaceXSI and X. Among them was a confidential onboarding plan for a member of the enterprise sales team, containing internal system addresses and 19 chat channels, as well as a sales territory map featuring 594 strategic accounts and pipeline information. Eight advertising proposals from X were also found, including one marked confidential that targeted a U.S. federal agency. Reports and dashboards of advertising expenses were available. The map distributed 594 accounts across six geographical areas based on their priority. Additionally, medical test results of patients, professional schedules, and unencrypted passwords were identified. Confirmed situations included a report from an ICU with a patient's test results, a list of over 3,500 employees from a hospital group, and a firefighters' schedule. The source code of a time management application made 18 accounts accessible along with unprotected passwords. Thirty affected organizations were identified, including six Fortune 500 companies, without disclosing their names.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Measured Scope: 35,578 Open Apps Out of 37,571 Verified

Claroc claims that nearly 95% of over 37,000 Grok applications are accessible without authentication and says it found over 35,000 open. Out of 37,571 applications still online that it verified, 35,578 were accessible, representing 94.7%. According to analyses, the majority correspond to games, homepages, or calculation tools, but nearly 10% of them contain confidential, internal, or personal information. Almost two percent present a high risk, and researchers estimate that 650 to 800 applications contain private data. Compromised information included patient medical records as well as confidential files from SpaceXSI.

Technical Causes Pointed Out in the Publication Flow

Claroc attributes these exposures to an unsecured default publication setting, easily discoverable addresses, and failing access controls in code generated by Grok. The company identifies the publication dialogue as the source of the problem and notes that, on the paid account used for its tests, the option "Anyone with the link" was pre-selected. The addresses were easily traceable because each application receives its own subdomain, referenced in passive DNS databases that can be queried by pattern after an initial visit. SpaceXSI's Build mode, previously named xAI, allows users to describe an application in chat and then let Grok write the code; an application becomes public when its creator clicks Publish, which hosts it on grok.me.

Tests Opening Writing and Deletion to Third Parties

Tests conducted by Claroc on its own applications revealed other failures. A contact list application ended up with an unconnected database, allowing anonymous visitors to add records. For a sales team client list, Grok automatically added Google and X login, but the generated code first added any authenticated user to the team before verifying their membership. An external Google account could thus read, modify, and delete data. Claroc specifies that it limited writing and deletion tests to its own applications.

Positions of SpaceXSI and Proposed Fixes by Claroc

SpaceXSI claims that applications remain private until shared by their creator, that publication requires several steps, including a legal notice, and that employee accounts are by default set to "Only you." Claroc responds that these steps do not prevent exposure if the "Anyone with the link" option remains pre-selected at the end of the process, citing the 24 open employee applications to illustrate the ineffectiveness of the stricter setting. The company recommends making "Only you" the default setting, displaying the recipient list before publication, and alerting if an application has a database without a login mechanism. It also suggests replacing individual subdomains with hard-to-guess paths on a shared host. Claroc indicates that it offered SpaceXSI to confidentially transmit the complete list of addresses and, lacking a commitment to notify the affected owners, is working to notify them itself.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.