Skip to content

Guides

Vibe coding security: Lessons from the Grok Build leak

Vibe coding security matters when AI builds your app. Learn what the Grok Build leak exposed and the checks to run before you publish.

On this page
  1. What happened in the Grok Build leak
    1. The short timeline
    2. What data was exposed
  2. How an app link became public
    1. Passive DNS in plain language
    2. Why unguessable links are not enough
  3. Are AI-built apps safe?
  4. A checklist before you publish any AI-built app
    1. How to stop app link exposure
    2. Secure no-code app links: a quick self-test
  5. Build it with FromScratch
  6. FAQ
Illustration of vibe coding security concerns, with a padlock in front of a laptop showing an app dashboard

Vibe coding security is easy to overlook when an app takes an afternoon to build. A recent incident involving Grok Build, the vibe-coding agent from SpaceXSI, shows how one design choice in how app links are created can expose internal material to anyone who finds the address. This article explains what happened, why it matters if you build without writing code, and what to check before you publish.

What happened in the Grok Build leak

According to Engadget, which cites Ars Technica, a flaw in SpaceXSI's Grok Build agent allowed sensitive data from thousands of users to become visible publicly. The security firm Claroc found the problem and reported it. Claroc's researchers said the exposure reached data from a hospital group, a fire department and SpaceXSI itself.

The issue was not a hack in the usual sense. Nobody broke into a server. Instead, app links generated by the tool could be picked up by outside systems and then opened by people who were never given them. Claroc called these pages "GhostShares," because they could be found and viewed without anyone sharing the link.

The short timeline

The sequence, as reported, went like this:

  1. Grok Build generated links for apps, and those links could be discovered through public internet records.
  2. Claroc documented the exposed apps and shared its findings with SpaceXSI and the affected companies.
  3. Some affected companies moved quickly to contain the exposure.
  4. SpaceXSI reportedly did not respond until Claroc contacted company leadership publicly on X.
  5. SpaceXSI then took down 24 of its own exposed apps. According to Claroc, the apps other Grok users and organizations left open had not been taken down.

The exact dates are not part of the reporting excerpt, so treat this as a sequence of events rather than a calendar.

What data was exposed

The material was not limited to public pages. Claroc's findings included an onboarding plan, a sales territory map covering 594 strategic accounts, eight X advertising proposals, four live advertiser account boards and two copies of an X brand safety playbook. Much of it was internal business material.

Only one of the exposed apps asked for a password. Claroc also reported that six of the affected organizations appeared on the Fortune 500 list and 26 had more than 1,000 employees. Its estimate was that about two percent of the exposed material was serious enough to rate high or critical.

The core of the problem was the link itself. Grok Build used "dice words," randomly generated phrases of four words, with up to 38 bits of entropy. That is a strong amount of randomness for a link. Guessing the phrase would be hard.

The weakness was placement. According to the reporting, Grok Build put the random part of the link to the left of the domain name. Google's Gemini, by comparison, places it to the right. That difference let the link show up in places it should not have.

Passive DNS in plain language

Every web address begins with a domain name, and computers look up that name through DNS, the system that translates names into server locations. Passive DNS services record those lookups as they happen across the internet and store them in searchable databases. If part of your app's address is written in a way that gets logged, anyone who searches those databases can find it, even if your app was never linked from a public page.

Many people think a long, random link is a form of protection. It is better than a short, obvious one, but it is not access control. Access control means the system checks who is asking and whether they are allowed to see the content. A random link only works as a secret if it never leaks, and this incident shows how easily a link can leak through logs, shared messages, browser history or search databases.

For anything sensitive, the link should open a page that asks for a login, or it should not open at all. A login check is a real control. A hard-to-guess URL is a convenience.

Are AI-built apps safe?

The honest answer is that they can be, but safety is not automatic. An AI builder creates the interface, the back end, the database and the links for you. That speed is useful, but it also means the default settings matter a great deal, and you may not see them unless you look.

Here is what this incident changes for people building without code:

  • Platform choices affect you. How a builder formats links, which pages are public by default and whether login is offered are design decisions made for you. They are worth asking about before you commit to a tool.
  • Publishing is your responsibility. When you publish, you decide who can reach the app. A public URL means the app is visible to anyone who finds it.
  • Internal tools are the most exposed. Dashboards, client portals and admin panels often hold employee records, customer details or financial numbers. Builders tend to share them with teammates through links, and those links are where things leak.
  • The impact is not limited to big names. The affected organizations in this case included a hospital group and a fire department, not only a technology company. A small business with a customer list faces the same risk.

If you want a broader view of the trade-offs that come with fast builds, the vibe coding mistakes to avoid before launch guide covers common errors in more detail.

A checklist before you publish any AI-built app

Use this list each time you put an app online, even a small one.

  • List what the app stores. Write down whether it holds personal data, employee records, customer messages, payment details or internal plans.
  • Decide who should see it. Name the people or groups who need access. If the answer is "anyone with the link," that is a risk for anything beyond a public brochure.
  • Turn on login before sharing any link. If the app holds anything sensitive, require sign-in on every page that shows real data.
  • Keep real data out of previews and tests. Use sample records while you build. Add real data only after access controls are in place.
  • Remove or retire old links. Delete apps you no longer use. Check that earlier versions are not still reachable at their old addresses.
  • Check what an outsider sees. Open your app address in a private browser window, signed out, and confirm what appears.
  • Keep a record. Maintain a simple list of every app you have published, its owner, its public address and whether it requires login.

Ask your builder these questions before you rely on it:

  • Are links to my apps public by default, or private?
  • Can I require login on individual pages or the whole app?
  • Are the addresses random in a way that keeps them out of public logs, or are they readable?
  • How do I remove access quickly if a link is shared by mistake?

If the answers are vague, treat that as information. Our guide on how to choose a no-code AI app builder includes a list of questions to compare platforms.

Pick one published app and do this:

  1. Copy its address.
  2. Open it in a private or incognito window, with no account signed in.
  3. Note what loads and whether any page asks for a login.
  4. Try a few obvious paths, such as pages for admin or export.
  5. If anything shows real data without a login, turn off public access and fix it before sharing the link again.

Build it with FromScratch

You can turn the checklist into a working habit by building a small internal register that tracks every app you publish. The steps below assume you have a FromScratch account. For background on how these tools work, see what an AI app builder is.

  1. Describe the register in plain language. Open a new project in the chat and explain what you want. The example below is a good starting point.
  2. Wait for the live preview. The agent builds the interface, the back end and the database, and the app appears in a preview within minutes.
  3. Test the form. Add a sample app, such as a client portal with a login requirement, and check that the list sorts the way you expect.
  4. Refine by chatting. If something is missing, ask for it directly, for example "add a filter for apps that hold customer data." Each change is saved, so you can return to an earlier version if a refinement goes wrong.
  5. Set the register itself to require login. Ask the builder to protect the register with sign-in before you add real entries. It holds your list of links, so it should not be public.
  6. Publish deliberately. You can publish to a public URL or a custom domain. Only make the register public if you have a specific reason, and keep it on a restricted address if you do not.

Type this prompt into the chat:

Build an internal tool called App Access Register. It should let me add each app I publish with its name, owner, public URL, purpose and a required field for 'Login required? yes/no'. Show a list sorted by apps with no login first, highlight any app that holds customer or employee data, and include a checkbox for 'reviewed this month'.

Once the register is running, review it at the same time each month. If you need help with setup or access questions, you can reach the team through contact and support.

FAQ

Not on its own. The Grok Build case shows that a strong random phrase can still leak if it is placed somewhere it gets logged or indexed. For anything sensitive, use authentication so that the page checks who is asking before showing data.

Does this mean AI app builders are unsafe?

Not automatically, but defaults and link design matter. Compare platforms by asking how they handle public access, login and link formats, and test your own apps the way an outsider would see them.

Turn off public access or take the app offline first. Then change or remove the link, review what was visible, and check whether any personal data was involved. If it was, follow the notification rules that apply to you and tell the people affected.

Do I need a developer to secure my app?

For basic access control, a clear checklist and a login requirement cover a lot of ground. More complex requirements, such as industry compliance or sensitive financial data, should be reviewed by a qualified security professional.

Written by

FromScratch TeamEditorial team

The team building FromScratch, the AI app builder that turns a plain-language description into a real, published full-stack app.