Every Name Is Listed

Researchers pulled about 43,600 Grok Build app hostnames out of passive DNS and found 35,578 apps open without a sign-in. A random name stops being a secret once it resolves, and the record that lists it isn't yours to delete.

Issue 41 - Nov. 17, 2026
By -top:1px solid #d4d4d8;font-size:0;line-height:0"> 

The News Voted Again

I closed last week's issue by promising tamper-evident records for this one unless the news voted otherwise; it did, two weeks running, and I've stopped pretending to be surprised. A security firm called Claroc published "GhostShares 2" on Nov. 10, studying apps people built with Grok Build, the in-chat app builder from the company behind Grok (Ars Technica's headline on the story calls it SpaceXSI; Claroc's own title still says xAI). Claroc's researchers pulled about 43,600 app hostnames out of passive DNS, loaded each app's first page once without signing in and found 35,578 of them open to anyone. A couple dozen were built by the company's own staff.

The news didn't wander far from the plan, though. Last week we set up two records, the proxy log for where data went and the transcript for what the model wrote; both are yours, kept as carefully or as carelessly as you like. This week's story runs on a third record, one you didn't know you were writing, and nobody can tamper with it or delete it (you included). Wherever a passive DNS sensor sits near a resolver, a lookup leaves a note that the name exists; the people who built these apps wrote themselves into that record the first time anyone opened a link.

My position is short. Once a name resolves in public DNS, treat it as listed, however many random words went into it. Who can reach what an agent built is a separate question from what the model saw, and it needs its own test; a vendor's default, or a vendor's server-side flag, answers that question for you until you check.

 

FOR FURTHER READING

 

What Claroc Found

Grok Build makes apps inside a chat on the web, iOS and Android; per the company's Aug. 19 post, it ran as an early beta for SuperGrok Heavy subscribers in July and opened to all plans that day. Each published app gets its own hostname under grok.me, often four random words long, and visibility can be set to "just you, anyone with the link, or public." The same post says apps show up as inline banners on X. The product was built to be passed around.

Claroc's method was modest, and the modesty is what should worry you. The team enumerated names by pattern from passive DNS (services that log hostnames as they're looked up, so a name opened once tends to stay findable), then requested each app's first page a single time, unauthenticated. Of the 37,571 apps still online, 35,578 opened without a sign-in, or 94.7%; the other 1,993 asked for a login. Nobody guessed a password or broke anything to get there.

Most of it was harmless. Games, calculators and ordinary websites make up the bulk, and Claroc estimates roughly 650 to 800 apps held private data. Per Claroc, that slice included an ICU patient's lab results, a hospital roster of more than 3,500 staff, a cruise crew list with IDs and a timesheet app storing 18 accounts' passwords in plain text (somebody's agent built a password manager by accident). Claroc also found internal data from 30 other organizations, 26 of them with more than 1,000 employees and six in the Fortune 500.

The staff apps made the headlines. Claroc counted 24 open apps built by SpaceXSI and X employees, including a sales-hire onboarding plan, a territory map covering 594 accounts, X ad proposals (one marked confidential and prepared for a U.S. federal agency) and campaign reports. Two access-control bugs round it out, per Claroc: a test app with no sign-in let any visitor change its data, and a team app's sign-in let an outside Google account read, edit and delete records.

 

Four Random Words and a Phone Book

The old phone book explains the mechanism better than any diagram. An unlisted number kept you out of the printed book, and it never once stopped a call from somebody who already had the number; the operator who connected that call still wrote it down. A four-word hostname is an unlisted number, and passive DNS is the operator's ledger. Random words protect a name from guessing; they do nothing once the name sits in a ledger other people can search.

I come at this from an odd angle, since part of my week goes to serving legal papers. In that work an unlisted number is a speed bump at most, because an address lands in a dozen records its owner never thinks about. Hostnames behave the same way. A name leaks into resolvers, browser histories, chat messages and any post that carries the link; the person who clicked publish saw none of it.

The 2025 stories looked similar and weren't; the gap matters for anyone reaching for a fix. On Aug. 1, 2025, OpenAI pulled ChatGPT's "Make this chat discoverable" option after Fast Company found shared chats in Google results; OpenAI CISO Dane Stuckey said "we think this feature introduced too many opportunities for folks to accidentally share things they didn't intend to." On Aug. 20, 9to5Mac, citing Forbes, reported that more than 370,000 Grok chats had been indexed through the share button. Those were crawler problems, and a noindex tag or a robots.txt rule speaks to a crawler. Passive DNS doesn't read robots.txt; the name lands in the ledger at lookup time, before any page gets fetched, so the 2025 fix has nothing to grab here.

 

Who Clicked Share

The users chose to share these apps, so this is a user problem. Right? Partly, and the record is disputed where it counts most. Claroc says its paid account came with "Anyone with the link" preselected; the company told a reporter, per TeslaNorth, that publishing takes several steps and that employee accounts default to "Only you." Both can be true at once (a paid account and an employee account aren't the same account), and nothing public settles it yet.

Either way, Claroc's vendor-side fixes point away from the user. It recommends serving apps under one host with unguessable paths, so a share link never becomes a DNS name, and making "Only you" the default. A path under one hostname travels inside TLS; a hostname gets looked up in the open. That's an architecture choice the person typing into a chat window never saw, and no amount of user caution changes it.

The timeline doesn't flatter anyone. Claroc reported to SpaceXSI on Nov. 8, reported 22 more staff apps on Nov. 9, heard nothing and then tagged Elon Musk and Grok manager Matt Palmer on X. Within hours of emailed questions from Ars Technica on Nov. 10, the company took the staff apps down; as of Claroc's post, other users' apps stayed up, and Claroc says SpaceXSI answered a reporter's questions while the report itself sat unanswered. Claroc then began contacting the affected organizations itself. Nobody has said when the exposure began; the product has been open to all plans since Aug. 19, which bounds the question and settles nothing.

 

The July Uploads

The July episode involved a different product, the Grok Build CLI coding agent, with the same lesson from the other side of the wire. Per The Hacker News and The Register, both on July 14, a researcher posting as cereblab found that the CLI (version 0.2.93, per THN) bundled whole Git repositories with full history and sent them to a Google Cloud Storage bucket named grok-code-session-traces. On a 12 GB repo, about 5.10 GiB went to storage while model traffic came to roughly 192 KB.

The details got worse from there. A canary file the agent was told not to open turned up in the bundle, and a tracked .env file went out unredacted. With "Improve the model" switched off, the uploads continued. The company's statement on X said zero-data-retention users store no code or trace data and that other users could run /privacy to disable retention and delete synced data; The Register reports that cereblab found /privacy wasn't what stopped the uploads.

What stopped them was a server-side flag, disablecodebaseupload set to true, flipped July 13 with no client update. THN cites analysis showing the upload code still present in build 0.2.99, gated only by that flag. Musk wrote on X that, as a precaution, all user data uploaded before then would be deleted; The Register said it couldn't independently verify the deletion. So the fix lived on the vendor's server, the off switch lived on the vendor's server and the code that did the uploading stayed on yours (an odd place to store a loaded thing).

My reading, labeled as mine: this is the case where last week's gate log earns its keep. TLS hides headers, which was the whole confession in #40, but it doesn't hide destination or volume; a proxy in front of that CLI would've logged a tunnel to a Google storage endpoint carrying something like 5 GiB against a model conversation measured in kilobytes. The meter from #37 was built to count, and this is what counting is for. Check that your log line records request size (in Squid that's the %>st logformat code; confirm on your version how it counts tunnel bytes), because a log that notes only the destination would've waved this through on any allowlist that already trusts Google storage.

 

Let's Call It Listed Engineering

To be LISTED, in the phone-book sense, is to be findable by anyone holding the book; the engineering is designing as if every name your agents produce is already printed in it. It has three parts, and none of them asks you to trust a random string.

First, names are listed on creation. A hostname in DNS, a certificate in the transparency logs or a link pasted into a chat is in somebody's ledger, so an agent-built artifact holding private data needs a sign-in that holds up whether or not the address is known. Second, reach gets its own test. What the model saw and who can open what it made are two questions with two answers, and the second one gets checked from outside: an account that isn't yours, a network that isn't yours and a browser with no session cookie. Claroc's outside-Google-account finding is that test, run by strangers instead of the owners.

Third, the vendor's defaults belong to the vendor until you check them. A preselected sharing option, a retention command that didn't govern the uploads and a flag on someone else's server each decided something about your data, and none of them was a control you held. I use vendor tools every day, and the July flag was a reasonable emergency fix; it's also a switch that can be flipped back without a release note, so I'd rather keep my own record of what left and my own test of what's reachable.

Assume the book is already printed.

 

QUICK TIP

Test Your Share Links From the Outside

Put the share links your team has published into a text file, one per line; this sends no cookies, so it sees what a stranger sees.

#!/usr/bin/env bash
# usage: ./reach.sh links.txt
while IFS= read -r url; do
  [ -z "$url" ] && continue
  code=$(curl -s -o /dev/null --max-time 15 -w '%{http_code}' "$url")
  case "$code" in
    200)         echo "OPEN      $url" ;;
    401|403)     echo "CLOSED    $code $url" ;;
    301|302|303|307|308) echo "REDIRECT  $code $url" ;;
    *)           echo "CHECK     $code $url" ;;
  esac
done < "${1:?usage: $0 links.txt}"

A redirect usually means a login page, and an app can serve its own login screen with a 200, so open every OPEN line by hand before you relax.

 

Quick Wins

🟢 Easy (~10 min): List the AI app builders your staff use, then run the script above on every share link you can find from them, starting with anything that touches customer or employee data.

🟡 Medium (~1 hour): Pull the names certificate transparency has already published for your own domains with curl -s 'https://crt.sh/?q=%25.example.com&output=json' | jq -r '.[].name_value' | sort -u, and ask whether each one should be reachable. It covers your domains, not a vendor's, and a builder that serves apps under a wildcard certificate won't show individual names there.

🔴 Advanced (half day): Add request size to your proxy's log line, then alert on any tunnel that sends more than a set number of megabytes to a storage host; tune the threshold on a week of normal traffic before you trust it.

 

Next Week

Next week, unless the news votes otherwise (it's two for two now), tamper-evident records finally get their slot. Passive DNS is the record nobody can edit. The transcript and the proxy log are records an agent or an attacker could edit; the July agents from #39 already tried.

Test every share link from an account that isn't yours.

Put a sign-in in front of anything an agent builds with real data in it.

About 43,600 names went into a ledger that neither the app owners nor the vendor controls, most of them for games and calculators and several hundred for things like patient labs and timesheet passwords; so treat every name your agents create as printed in the book, keep the lock on the door rather than the secret in the number, and check the vendor's switches yourself, because the ledger keeps its copy whether or not the app comes down.

Bobby R. Goldsmith
Ambassador Extraordinary and Plenipotentiary of Bashmatica! by NodeBridge Automation Solutions

P.S. If someone forwarded this, subscribe at bashmatica.com (it's the one hostname I'd like printed in every ledger going).

NODEBRIDGE AUTOMATION SOLUTIONS

Every guardrail in this newsletter, already wired into your Claude Code setup.

NodeBridge configures Claude Code against your actual repos: MCP servers, subagents, hooks, permissions and a project memory that keeps the right context in and the stale context out. You own every config when it ships. Setups start at $750 and go live in about three days.

See the Setup Packages