Breakdown: 35,578 Grok-built apps opened for anyone, because "anyone with the link" was the default and the links were in DNS
By Abbi Labs · Published 22 November 2026
What does a stranger see in your app?
The Abbi Labs quick check opens your live app as a signed-out visitor and looks for secret-looking keys in its JavaScript, exposed .env and .git files, and Supabase tables readable with your public key. It can't tell whether a page that opens was meant to be private, and it doesn't try signing in with a stranger's account, so do the checks below too. No sign-up.
No app handy?
No sign-up. Read-only public-resource and protocol metadata checks. No model prompts or tool execution. Scan only sites you own or are allowed to test.
Run on every deploy or use from Cursor / Claude Code via MCP →
Weekly AI-app security checklist
One short email a week: a practical check for apps built with Supabase, Next.js and AI tools (RLS, leaked keys, headers and more). Unsubscribe anytime.
Most AI app builders now end the same way: you describe the app, it works, you click Publish and get a link to share. On 10 November 2026, security firm Claroc published research on what happened to the apps people publish from Grok's chat-based builder. Of 37,571 Grok apps it found still online under grok.me, 35,578 (94.7%) opened for anyone, with no sign-in. Twenty-four of them were built by staff at SpaceXSI, the company that makes Grok, and at X.
Nobody needed to guess a link. Each app's random name is its own hostname, and hostnames end up in passive DNS databases that anyone with an account can search. Everything below comes from Claroc's report, with The Hacker News and Ars Technica as cross-checks. I'm not naming any of the other organisations or apps. Claroc didn't either.
What happened
- The publish step defaulted to open. Nothing goes public until the builder clicks Publish. But in the paid account Claroc used, the access step that follows offered "Anyone with the link", "Invited users" and "Only you", with "Anyone with the link" already selected.
- The link is a hostname, so it doesn't stay secret. Every published app gets its own subdomain of
grok.me. Grok suggests four random words as the name, and about half the names Claroc saw followed that pattern. Hard to guess, but the first time anyone opens the app, their computer looks the name up in DNS, and passive DNS services record those lookups. Claroc listed 43,601 names undergrok.meusing common subdomain discovery methods. - Claroc opened each first page once, signed out. It didn't sign in, submit forms, call server functions or use any password or key it saw.
- The sign-in Grok added didn't check who belonged. Claroc built two test apps with made-up data. A contact list got a database and no sign-in, and a visitor in a private window could add a row the next visitor saw. A customer list "for my company's sales team of 4 people" got "Continue with Google" and "Continue with X" buttons, but the code added whoever signed in to the team before checking whether they were on it. An uninvited Google account viewed every customer, edited one record and deleted another. Claroc ran these tests only on its own apps.
- The right settings exist. Switched to "Only you", Claroc's test app returned 401 to a stranger for both the page and its data, and unpublishing took an app offline within seconds.
The numbers
- 35,578 of 37,571 live Grok apps (94.7%) opened with no sign-in. 1,993 returned 401.
- About 1 in 10 open apps held internal business or personal data, roughly 650 to 800 apps by Claroc's estimate. About 2% held material Claroc rated high or critical.
- 24 open apps built by SpaceXSI and X staff, including a confidential onboarding plan, a sales territory map of 594 accounts and an X advertising proposal to a U.S. federal agency marked confidential.
- 30 other organisations identifiable from their open apps. 26 have more than 1,000 employees, and 6 are in the Fortune 500.
- Examples Claroc confirmed by hand include a hospital group's roster of more than 3,500 employees, an intensive care briefing on one patient, a fire department's live shift board, and a contractor's timesheet whose script file held 18 accounts with plain-text passwords.
What's confirmed and what we don't know
- Confirmed by Claroc: the counts above, the preselected "Anyone with the link" in its paid account, and the two test apps.
- SpaceXSI's position: it told a reporter that Grok apps stay private until their creator shares one, that publishing takes several steps including a legal disclaimer, and that employee accounts default to "Only you". It took down the 24 staff apps after the report. Claroc says it has not done the same for other users' open apps, and its response didn't say it would contact their owners.
- Not public: whether anyone else found or misused the open apps. Claroc only tested writes on its own two apps, so we don't know how many of the 35,578 would also accept changes.
- Not just Grok. Claroc notes that Lovable, Replit, Base44, Google AI Studio, ChatGPT and Meta's Muse also give each published app its own hostname, and v0 and Bolt deploy to hostnames too. ChatGPT's builder defaulted to "Only me" in Claroc's test. The difference is the default, not the DNS.
Check the apps you've published
This works for any AI builder, not just Grok. Do it for every app you've ever shared a link to.
- Make a list. Every app you've published from Grok, Lovable, Replit, v0, Bolt, Base44, Google AI Studio or ChatGPT, including the "just for the team" ones and the old experiments. If nobody owns it any more, unpublish it.
- Open each one signed out. Use a private window, or from a terminal:
A private app should give you an error page (Claroc saw 401 for "Only you" on Grok). If you getcurl -s -o /dev/null -w "%{http_code}\n" https://your-app.example/200and the real app, so does everyone else. - Assume the link is public. If a hostname has been opened even once, it can be in a passive DNS list. "Anyone with the link" means anyone. For anything internal, pick "Invited users" or "Only you".
- Check the data, not just the page. Open DevTools, go to Network, reload while signed out, and look at the data requests the page makes. Real records coming back without a sign-in means the data is open even if the page looks empty.
- Try a stranger's account. Sign in with a second Google account that isn't on the team. If it gets in, the app checks who you are but not whether you belong. The fix is an allowlist the owner controls, checked before granting anything. Illustrative, not Grok's actual code:
// wrong: anyone who signs in is added first, so the check always passes await addToTeam(user.email); if (!isTeamMember(user.email)) return deny(); // right: check a list the owner controls before granting anything const ALLOWED = new Set(["you@yourco.com", "teammate@yourco.com"]); if (!ALLOWED.has(user.email.toLowerCase())) return deny(); - Search the scripts for things that shouldn't be there. Everything in your JavaScript is downloaded by every visitor:
Names, emails or passwords in there are public. Move them to a database behind sign-in, and change any password that was ever in the code.curl -s https://your-app.example/ | grep -oE 'src="[^"]+\.js"' curl -s https://your-app.example/assets/index-XXXX.js \ | grep -oiE '(password|passwd|secret|api[_-]?key|@[a-z0-9-]+\.(com|org))[^,;]{0,40}' | head - No sign-in plus a database means open writes. Claroc's test contact list let any visitor add rows. If an app stores data, either require sign-in or make it read-only for visitors.
Check what a signed-out visitor gets
Paste your app's URL and Abbi Labs looks at it the way a stranger would: secret-looking keys in your JavaScript, exposed .env and .git, open Supabase tables. Read-only. It won't know whether your roster was meant to be private, so steps 2 and 5 above are still on you.
If you find an open app with real data
- Change the access setting or unpublish it first. On Grok, Claroc saw unpublishing take an app offline within seconds.
- Change any password or key that sat in the app's code. Hiding the app doesn't un-copy what was already downloaded.
- Work out what was in it and for how long. If it held personal data about customers, staff or patients, talk to whoever handles privacy at your organisation. That may be a notifiable breach where you are.
- Tell your security team about the builders people use. Claroc's point is that these tools need no IT rollout, so nobody else knows the apps exist.
My take
No one hacked anything here. The apps did what their settings said. The problem is a default that treats a random-looking name as a lock, when that name gets written into DNS the first time someone uses it. A sales map and a lunch poll get the same Publish button, and the person clicking it usually isn't thinking about passive DNS. Open every app you've published in a private window today. It takes a minute per app.
Sources
- Claroc Research, GhostShares 2: Grok Apps Exposed xAI's Business Secrets and People's Private Data, 10 November 2026 (primary; counts, test apps, disclosure timeline and SpaceXSI's response)
- The Hacker News, Over 35,000 Grok-Built Apps Exposed Online, Some Leaking Confidential Data and Passwords, 11 November 2026
- Ars Technica, Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI, 10 November 2026 (company statement; staff apps removed after Ars asked)
Related
- Breakdown: RedAccess found 5,000+ vibe-coded apps with almost no authentication
- Breakdown: a vibe-coded agent dashboard at METR failed open and leaked an API key worth $600k in credits
- Breakdown: Crissy Field found thousands of Lovable apps with open Supabase tables
- OpenAI, Claude or Gemini API key exposed in your frontend? What to do now
- Supabase "authenticated users only" policies: why any stranger who signs up passes them
- Exposed .env, .git and source maps