📑 Table of contents

Skynet Watch: 94.7% of apps created with Grok Build were accessible without an account, including 24 built by teams from SpaceX and X

Skynet Watch 🟢 Beginner ⏱️ 14 min read 📅 2026-11-13

Skynet Watch: 94.7% of apps created with Grok Build were accessible without an account, including 24 built by SpaceX and X teams

🔎 Nobody had to force anything

On November 8, 2026, researchers at Claroc loaded 37,571 applications published via Grok Build, xAI's app-building agent. 35,578 opened without any credentials whatsoever. That's 94.7% of the catalog, readable by anyone who got their hands on the link — including 24 applications built by SpaceXSI and X employees, among them a confidential advertising proposal intended for a US federal agency.

No exploit, no exfiltration, no zero-day. Grok Build's publishing panel pre-selected "Anyone with the link": everyone with the link. The SpaceXSI staff apps went through the full publishing process, legal disclaimers included, only to land in open access on grok.me subdomains that passive DNS lists like a directory.

That's the entire lesson of this affair, and it goes far beyond xAI: when agentic building tools drive the cost of publishing down to zero, they drive the cost of misconfiguration down to zero in the same stroke. The default becomes the de facto standard, and nobody re-reads the checkbox the tool ticked on their behalf.

The same week, Notion confirmed that a breach at Fireworks, its AI infrastructure provider, had exposed Notion AI chat transcripts — less than 1% of workspaces, but the mechanism is the same: your data flows through a chain of third parties you don't control, and the leak always ends up at the weakest link.


Key takeaways

  • 37,571 Grok Build apps tested online, 35,578 accessible without logging in (94.7%), according to the research brief from the Cloud Security Alliance (November 2026).
  • About 1 in 10 open apps contained internal or personal data; ~2% of the material rated high/critical, an estimated 650 to 800 apps.
  • 24 apps from SpaceXSI and X staff: confidential onboarding plan, sales territory map covering 594 accounts, advertiser decks, ad proposal for a US federal agency.
  • Systemic cause: per-app grok.me subdomains, indexed via passive DNS, and the "Anyone with the link" default preselected in the publishing panel.
  • Identical precedent one month earlier: Meta Muse, 6,800 hostnames exposed through the same mechanism.
  • ChatGPT Apps uses the same per-app subdomain architecture, but with an "Only me" default — proof that the problem isn't the architecture, it's the setting.
  • SpaceXSI pulled its 24 apps within hours of an email from Ars Technica; the rest of the catalog was still online when the GhostShares 2 report was published on November 10.

Whether you're auditing your current exposure or starting over on a well-controlled foundation, four tools cover the bulk of the work.

Tool Primary use Price (November 2026) Ideal for
Censys Inventory of exposed hostnames and services Free, pro plans by quote (check censys.io) Security teams mapping their exposure surface
Shodan Detecting publicly accessible apps and services Free, subscriptions from a few dollars (check shodan.io) One-off audits and monitoring of a domain
SecurityTrails Passive DNS and subdomain history Limited free tier, paid plans (check securitytrails.com) Finding your forgotten grok.me subdomains
Hostinger VPS to self-host your internal tools with real authentication From ~€2.99/month (November 2026, check hostinger.com) Teams that want to keep control of their data

The exact numbers: 37,571 apps tested, 35,578 open

Out of 37,571 Grok Build applications online, 35,578 — or 94.7% — loaded without any authentication. That's the core of GhostShares 2, the report published by Claroc on November 10, 2026, after a discovery on November 8, detailed in a note from the Cloud Security Alliance.

The breakdown is more worrying than the overall figure. Roughly one in ten open apps — about 3,500 — contained internal or personal data. And about 2% of the exposed material was rated high/critical, bringing the estimate to 650 to 800 applications containing genuinely sensitive information.

An important methodological clarification: Claroc didn't "break" anything. The collective enumerated subdomains via passive DNS, then loaded the pages. The entirety of the "hacking" fit inside a private browsing window.

To visualize what 94.7% means, imagine a parking lot with 37,571 cars, 35,578 of which have the key left in the ignition. The owner will tell you that "you had to know the spot." That's true. It's also completely irrelevant in the face of a robot that scans license plates.


"Security by Secret URL" Isn't Security

Grok Build links relied on random three-word names — roughly 38 bits of entropy — under grok.me subdomains dedicated to each app. Ars Technica analyzed this capability-based scheme and its blind spot: it works as long as nobody enumerates the hostnames. But passive DNS does exactly that, at scale and with archival persistence.

Passive DNS, in short, is the systematic recording of DNS resolutions observed on the network. Every subdomain visited by any visitor leaves a trace, and those traces aggregate into commercial databases. Your "secret links" are therefore already in someone's catalog.

38 bits is serious against a human guessing. It's trivial against a script reading a DNS zone. As soon as your subdomains are indexed, the "capability" becomes a public directory.

This isn't an isolated accident, it's a recurrence. The previous month, the same method revealed 6,800 exposed hostnames on the Meta Muse side. And ChatGPT Apps uses the same per-app subdomain architecture — with an "Only me" flaw. In other words: the technique is identical, the configuration makes all the difference.

My take is simple: a URL is not an access control, it's an identifier. Confusing the two means repeating the mistake of "public by link" sharing — only worse, because here anyone can index your link without even paying you a visit.


What leaked: a hospital directory, plaintext passwords, a sales plan covering 594 accounts

Roughly 1 in 10 exposed apps contained internal or personal data, and the top of the pack is severe. Three examples flagged by Claroc give a sense of the problem's scale.

A hospital directory listing over 3,500 people. A subcontractor's timesheet, with 18 accounts and their plaintext passwords — including an admin account. And on the SpaceXSI/X side, 24 apps: a confidential onboarding plan, a sales territory map covering 594 accounts, media plans and advertiser decks with budget tiers, and that now-famous confidential ad proposal for a US federal agency (TechRadar).

According to The Hacker News, 30 other identifiable organizations are affected — construction, energy, healthcare, defense — including 26 with more than 1,000 employees and 6 from the Fortune 500. Damning detail: only one of the apps tested by the researchers required a password.

One last detail says everything about the quality of the generated code. Of the two apps the researchers created to test the platform, one displayed an editable contact list with no login required, and the other — a client list — added anyone signing in with Google to the team before even verifying whether they belonged there. The agent generates code, not security: the authorization logic was inverted, and no one reviewed it.


Why SpaceX and X Teams Paste Sensitive Data into a Builder

Because it's the shortest path between a business need and a working tool — and because nobody audits the sharing settings of a tool that isn't theirs.

These builders are a straight line to productivity. A deck becomes an interactive app in minutes, a sales spreadsheet becomes a shareable dashboard with no IT ticket and no code review. SpaceXSI's territory map of 594 accounts was very likely built in a single afternoon by someone just trying to do their job.

This is shadow IT 2.0. Yesterday, the employee installed unauthorized SaaS; today, they generate an unauthorized application — one that contains real data. The security team has no inventory, no alerts, no visibility into this distributed catalog.

The marketing pitch of these tools — "from idea to app in one minute" — describes the problem precisely. One minute is also all it takes to publish a directory of 3,500 employees without realizing it. The friction that was removed was also the friction that protected.

Grok Build, the CLI coding agent launched by xAI amid a coding agents war where OpenAI's GPT-5.3 Codex and Anthropic's Claude Sonnet 4.6 already arm the competition, made publishing speed its central selling point. Security remained an afterthought.


SpaceXSI says no, Claroc shows the numbers, and the catalog stays online

SpaceXSI claims its apps were private by default and that nothing becomes public without an explicit choice from the creator. That's the position conveyed to Ars Technica through a spokesperson for Cursor — the studio acquired in August — who describes a multi-step publication process, with a legal disclaimer and an "Only you" default on employee accounts.

Claroc responds with a number: the 35,578 exposed apps made it through all those steps anyway. Fixed default or not, multi-step process or not, the result was there, measurable, and its name is 94.7%.

The company's reaction was swift within its own perimeter. SpaceXSI removed its 24 apps within hours of Ars Technica's email. The rest of the catalog remained online at the time of the report's publication. A notable fact: SpaceXSI asked for its apps to stay online once the default was fixed — a sign that usage still trumps caution, even among the exposed.

There is a gap between "private by default" as marketing tells it and as experience measures it. The first version is a product journey. The second is a test in private browsing mode. Only the second one counts.


The same week, Fireworks and Notion: the pattern of data leaking through AI

On November 13, Notion confirmed that a breach at its AI infrastructure provider, Fireworks AI, had exposed Notion AI chat transcripts in a small number of workspaces — less than 1% according to the company (BleepingComputer). Fireworks notified Notion as early as the evening of November 2; some chat sessions, which could not be attributed, may have been exposed.

The scale bears no comparison to Grok Build, but the mechanism is the same: your data leaves your perimeter, passes through a chain of providers, and the leak happens at a link you don't control. Less than 1% of workspaces — that tells you one thing: even the providers don't know exactly which ones.

This pattern isn't an isolated textbook case; it's the background noise of 2026. The same period saw a Russian network of voice clones rack up 28 million views on the French riots and ChatGPT Desktop silently transferring your SSH agent to remote hosts. In all three cases, AI expands the exposure surface faster than organizations can map it.

The daily digests (Byobot, The Neuron) happened to cover these stories on the same day, incidentally. That's not an editorial coincidence: it's the normal rhythm of an ecosystem where data moves faster than governance.


Check your exposure in 30 minutes — then fix it

Three checks are enough to know whether your apps — or your organization's — are exposed, and two settings are enough to fix it. The recommendations come directly from The Hacker News.

  1. Search for builder domains in passive DNS. SecurityTrails or Censys enumerate a zone's subdomains far more effectively than your memory.
  2. Test in a private browsing window. If the app loads without a login from another session, it's public to the entire world.
  3. Read the embedded scripts. Names, emails, passwords: generated code often hardcodes them into the page.

A minimalist command-line test:

curl -sI https://mon-app.grok.me | head -n 1

A 200 with no cookie or session means a public page. Period.

To fix it: revoke sensitive apps, change the sharing default, and republish behind real authentication — Cloudflare Access does this for free for up to 50 users (as of November 2026, verify on cloudflare.com). If you want to keep control end to end, self-host: a Hostinger VPS combined with Cloudflare Tunnel to expose your services without opening ports offers the same convenience as a builder, but with your own access rules and your own logs.

Two useful additions. If you manage a large AI-generated document corpus, stop searching blindly: vector-free RAG approaches that reason instead of searching help you map what you've actually produced and exposed. And to prototype without depending on a proprietary platform, a free model via OpenRouter or Groq on your own hosting does the job without handing your data over to someone else's default.


❌ Common Mistakes

The URL is not a secret, it's an address. Passive DNS indexes hostnames, and 38 bits of entropy weigh nothing against automated enumeration. Solution: real authentication, "Only me" setting, and a test in private browsing mode before pasting any data.

Mistake 2: prototyping with real data

The hospital directory and the timesheet with passwords in plain text only existed because someone pasted real data into a prototyping tool. Solution: synthetic data during the testing phase, and switching to an authenticated environment before any real content.

Mistake 3: trusting the tool's default setting

"Anyone with the link" being preselected is not a security decision, it's a product growth decision. ChatGPT Apps proves that an "Only me" default is technically identical to implement. Solution: audit the sharing settings of every app before putting anything into it.

Mistake 4: forgetting that generated code can have broken authorization logic

The customer list that added every Google signatory to the team before verifying their membership — that's an agent's code, not an auditor's. Solution: review the generated authentication code and systematically test with two accounts holding different roles.


❓ Frequently Asked Questions

What exactly is Grok Build?

xAI's app-building agent, launched as a CLI right as the coding agent war intensified. From a single prompt, it generates a complete web application powered by Grok 4.1 and publishes it to a dedicated grok.me subdomain. It's this subdomain-based deployment model that made mass enumeration possible.

Are the exposed apps still online?

SpaceXSI removed its 24 applications within hours of Ars Technica's report, but the rest of the catalog was still online when the report was published on November 10. xAI has since fixed the flaw, but any app published before the fix warrants an immediate audit.

How does it differ from October's Meta Muse leak?

The mechanism is identical: hostnames enumerated through passive DNS, open-by-default hosting. Meta Muse involved 6,800 hostnames; GhostShares 2 documents 35,578 apps on the Grok Build side. Two incidents in two months using the same method: this is an industrial pattern, not an isolated accident.

Is ChatGPT Apps safer?

Its per-app subdomain architecture is the same, but the sharing default is set to "Only me", which eliminates the mass exposure seen with Grok Build. It's not immunity — a wrong click is still possible — but the default setting provides protection where it counts.

I published an app containing internal data — what should I do?

Revoke or delete the app, rotate any credentials that were embedded in it, and assume the data has been indexed: passive DNS archives persist after deletion. Then inspect the embedded scripts for names, emails, and passwords left hardcoded in plain text.


✅ Conclusion

When the cost of publishing tends toward zero, the cost of a misconfiguration tends toward zero too — and 94.7% of Grok Build apps opening without an account is what that looks like in practice. Before pasting so much as a single piece of data into an agentic builder, run the test in a private browsing window: it's the only audit that never lies.