A shareable link only protects you if nobody can find it, and Grok Build's links could be found. If an agent in your stack can publish anything to a URL, assume someone is already scanning for it.
Grok Build, the vibe-coding agent from Elon Musk's SpaceXSI (the company formerly known as xAI), generated public links to the apps people built with it, and security firm Claroc found that anyone watching DNS records could discover them, Engadget reported on November 11. Ars Technica's coverage, as summarized by Techmeme, put the number of exposed apps at more than 35,000.
SpaceXSI's own apps were among them. According to Engadget, Claroc found 24 apps built inside the company that were reachable through public links, holding onboarding materials, a sales territory map covering 594 strategic accounts, eight X advertising proposals, four live advertiser account boards and two X brand safety playbooks. Claroc also counted six Fortune 500 companies and 26 organizations with more than 1,000 employees among the affected users, and SC Media's brief on the research mentions employee information from a hospital group and a fire department.
The links weren't guessable in the usual sense. Grok Build built them from 'dice words', random four-word phrases worth up to 38 bits of entropy, which on paper is hard to brute-force. The mistake was where the phrase went. Engadget says it sat to the left of the domain, as a subdomain, instead of in the path after it. A subdomain shows up in DNS lookups, and passive DNS services log those lookups into searchable databases. A path stays inside the encrypted request and never gets that exposure.
Claroc named the problem GhostShares. Its estimate is that about 2% of the exposed material was critical-level sensitive, which sounds small until you multiply it by tens of thousands of apps. Only one of the exposed apps asked for a password.
Grok Build Quietly Uploaded Developers Entire Codebases to xAI Cloud A developer's reverse-engineering of xAI's Grok Build tool found it silently up entire codebases, including deleted secrets, to xAI's cloud storage, directly contradicting the company's zero data retention claims. Elon Musk has promised to delete the data, but xAI has offered no timeline or proof. - grok build uploads entire codebases - xai cloud security breach developers
Frankly, a long random link was never access control. It's a convenience that works until the link leaks somewhere you didn't plan for, and here the leak was built into how the address was formed.
The response was uneven. Engadget reports that SpaceXSI took down its own 24 apps only after Claroc tagged Musk and another company official on X, and that the apps belonging to other organizations were not secured. We couldn't find a public SpaceXSI statement addressing those customer apps in the coverage we reviewed, so treat that part as unanswered.
It's also not the first Grok Build scare. In July, a researcher going by Cereblab found the tool up whole Git repositories, history and committed secrets included, to a Google Cloud Storage bucket run by xAI, as The Next Web and The Register reported at the time. Musk promised to delete the uploaded user data, and the company open-sourced the tool. Two data-handling problems in four months is a pattern you should price in.
What to check before an agent touches your code #
Start with the assumption that anything an agent publishes is public. If your team uses Grok Build, Claude Code, Cursor or any tool that can host or share what it makes, find out whether shared apps sit behind a login by default. In Claroc's sample, one did.
Then look at what the agent can reach. Give it scoped, short-lived credentials rather than your personal keys, and keep production data, customer lists and secrets out of anything it builds as a prototype. The SpaceXSI leak wasn't source code. It was sales maps, advertiser boards and onboarding decks, the kind of material people paste into a quick internal tool without thinking of it as a repo.
Keep an inventory of what your agents have published, with a named owner, so you can take things down within hours instead of waiting for a researcher to tag your CEO. Put publishing and sharing actions behind human approval, the same way you would a payment or a deploy. And if an agent vendor can't tell you how its share links are generated and indexed, that's your answer about how much to trust it with real data.
For now, Claroc's tally stands: one password-protected app among the exposed set, and 24 of them belonged to the company that built the tool.
[xAI brings Grok Build into the coding agent fight](https://startupfortune.com/xai-brings-grok-build-into-the-coding-agent-fight/)
xAI has launched Grok Build, a terminal-based coding agent available first to SuperGrok Heavy subscribers. The beta shows how coding agents are becoming a paid workflow battle, not just a model benchmark contest. - Grok Build coding agent for developers - xAI competing with Claude Code and Cursor
Also read: SAP keeps promoting its Joule AI tools but still hasn't published numbers to test them • Airbnb says its own AI travel agent will book trips on and off its platform • How to set an AI agent human in the loop approval threshold
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.