Gangsta AI

By Gordon Ramsay · 2026-11-11 · 4 min read
Satire. Parody written in the voice of Gordon Ramsay, who did not write or endorse it. About this column
Somebody at SpaceXSI built an app with Grok to onboard a new enterprise sales hire. It was marked confidential. Anyone who knew where to look could open it, along with a sales territory map covering 594 accounts and an X advertising proposal to a US federal agency.
Those were just the in-house dishes. Security firm Claroc says it found 37,571 apps built with Grok Build live on the internet, and 35,578 of them opened for anyone, with no sign-in. That is roughly 95% of the kitchen with the back door propped open.
What was sitting on the counter
Most of the open apps were harmless: games, landing pages, calculators. But Claroc says about 1 in 10 held internal business or personal data, and about 2% were serious enough to rate high or critical. The researchers' list, as reported by The Hacker News and Ars Technica:
- A hospital group's searchable list of more than 3,500 employees
- An intensive care briefing on a single patient
- A fire department's live shift board
- A contractor's timesheet whose code held every employee's password in plain text
- Galley checklists from a celebrity cruise line (the only part of this I am qualified to inspect, and even those were left out on the pass)
Beyond SpaceXSI and X, the open apps exposed internal data from 30 other organizations Claroc could identify. Twenty-six have more than 1,000 employees. Six are in the Fortune 500.
“The random four-word app names didn't help. Opening an app makes your computer look it up in DNS, and those lookups end up in searchable lists.”
How you serve 35,000 apps raw
Every published Grok app gets its own subdomain on grok.me. When Claroc's researchers clicked Publish on a paid account, the access setting was already set to "Anyone with the link." Passive DNS services record subdomain lookups, so the "secret" link isn't secret for long. Ars reports a passive DNS search returned 43,601 grok.me links.
Then it gets worse. Claroc built a test app for a four-person sales team, and Grok added "Continue with Google" and "Continue with X" sign-in buttons on its own. The code Grok wrote added each new user to the team before checking whether they belonged there, so the check always passed. A stranger's Google account walked in, viewed every customer, edited one record and deleted another. Claroc ran that test only on its own apps.
A representative for Cursor, which SpaceXSI acquired, told Ars that Grok Build apps "are private by default and nothing goes public unless the creator chooses to share it." Claroc says SpaceXSI pulled its own 24 staff apps but had not said it would warn the thousands of other owners. Claroc says it has begun emailing affected organizations itself.
For contrast: Claude artifacts, ChatGPT shared chats and Gemini shares keep the random part of the link in the URL path, so DNS never sees it. ChatGPT's app builder defaulted to "Only me."
The lesson: never let one cook taste his own soup
Look at what actually failed here. One model wrote the sign-in code. Nobody checked it. Not a second engineer and not a second model. The bug, "add the user, then check if they're allowed," is the kind of thing a different model often catches on sight.
And it isn't a Grok-only problem. Claroc says Lovable, Bolt and Replit apps get their own web addresses too, so the same DNS records expose them. With Claude Opus 5.5, GPT-6 Astra, Gemini and Grok 4.5 trading the lead every few weeks, no single model is reliably the one that gets it right, and none of them should be grading its own homework.
That's the case for Gangsta AI: ask the same question, whether it's "is this login check safe?" or "who can see this app?", across ChatGPT, Claude, Gemini, Grok and 30+ more at once. You get one cross-checked verdict instead of one confident guess. If four models flag the same hole, you fix it before you publish.
See which models are actually earning their stars right now on our best AI rankings. Then set your Grok apps to "Only you." This kitchen is closed.
Sources / Receipts
- Ars Technica — Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI
- The Hacker News — Over 35,000 Grok-Built Apps Exposed Online, Some Leaking Confidential Data and Passwords
- Claroc — research report on open Grok Build apps
- Engadget — Grok Build leaked sensitive SpaceXSI data
- Hero photo: SpaceX Starbase launch facility, Boca Chica, Texas — Lars Plougmann, Wikimedia Commons (CC BY-SA 2.0)
- Gordon Ramsay photo: Dave Pullig, Wikimedia Commons (CC BY 2.0)
← Newer: Apple Sues OpenAI, and the Altman–Musk Beef Goes Nuclear Again · Older →: Anthropic Wants $1.4 Trillion From the Public Market. It Lost $104 Billion in the Last Nine Months.
Related reading: Developers Clicked 'Don't Ask Again.' Claude Code Saved Their Passwords — and 1,122 of Those Files Are Sitting on Public GitHub. · OpenAI's Agents Built 948 Working Linux Root Exploits in One Run. Each One Cost $1,279. · A Teenager Pointed 40 Machines at the Internet's Biggest Fanfiction Archive While It Was Offline. In 24 Hours: 126,592 Visitors, 11 Audiobooks Sold, and 2,000 Takedown Notices. · Y Combinator Gave 50 AI Agents One Job: Build a Civilization. Four of Them Spent 11 Minutes Appointing a 'Sole Fertilizer Operator.'
More: AI leaderboard: best AI models right now · Compare AI models side by side · Best AI for coding · Best AI for image analysis · Frontier Models · All articles · Archive