Shadow AI: when employees ship apps, how to find them and what to test

Shadow AI is the use of AI tools at work without the approval or oversight of IT. Its first form was data pasted into a chatbot. Its new form is an employee who builds an app with Lovable, Cursor, Claude Code or Replit and publishes it with company data behind it. Blocking the tools does not remove those apps. Find them through hosting accounts, sign-in grants and certificate logs, then test each one from outside.

The team behind Keybrum's attacksUpdated 11 min read

In short

  • Shadow AI now has two forms. In the first, data leaves through a prompt. In the second, an employee publishes an app with a database, a login and API keys, and nobody in IT knows its address.
  • Blocking fails on numbers alone: in Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 knowledge workers in 31 markets, 78% of AI users said they bring their own AI tools to work.
  • Three records already hold most of the inventory: the project lists of your hosting accounts, the third-party apps your people signed in to with their work account, and the public certificate logs for your domains.
  • An app on a personal account under the platform's own domain appears in none of those records. Only an amnesty and a clear rule bring those in.
  • For each app found, answer five questions from outside: what a visitor with no account can read, which keys the JavaScript carries, who can sign up, whether one user can read another's records, and who can call the AI route.
On this page

What is shadow AI?

IBM defines shadow AI as "the unsanctioned use of any artificial intelligence (AI) tool or application by employees or end users without the formal approval or oversight of the information technology (IT) department". It is the AI chapter of shadow IT: people adopt a tool because it helps them today, and the company learns about it later, if at all.

The scale is not in doubt. Microsoft and LinkedIn's 2024 Work Trend Index, a survey run by Edelman Data and Intelligence among 31,000 knowledge workers in 31 markets between 15 February and 28 March 2024, found that 78% of AI users bring their own AI tools to work. More recent polls give different figures because they ask different questions, so treat any single percentage with care. The direction is the same in all of them.

It also has a cost. IBM's Cost of a Data Breach Report 2025, conducted by Ponemon Institute on breaches at 600 organizations between March 2024 and February 2025, reports that one in five organizations had a breach due to shadow AI, and that only 37% have policies to manage AI or detect shadow AI. Organizations with high levels of shadow AI saw breach costs 670,000 dollars higher on average than those with little or none.

The new form: employees who publish apps

Most writing on shadow AI stops at the chatbot: someone pastes a contract or a customer list into a prompt. That risk is real, and data loss tools were built for it. The second form is newer and those tools do not see it.

A sales manager describes a quote calculator to Lovable. A data analyst asks Claude Code for a dashboard on the customer export. An operations lead builds a supplier portal on Replit. Each of them gets a real app in an afternoon, with a database, a login page and a public address. They did nothing malicious. They solved a problem their IT backlog would have taken a quarter to reach.

Shadow AI, first formShadow AI, second form
What happensAn employee sends company data to an AI toolAn employee publishes an app or an agent built with an AI tool
Where the data sitsWith the AI vendorIn a database the employee created, reachable from the internet
Who can reach itThe vendor, under its termsAnyone who finds the address, if the access rules are missing
How long it lastsOne promptUntil someone deletes the app, often long after its author moved on
What sees itProxy logs, data loss tools, browser controlsAlmost nothing inside the company network: the app lives outside it

The apps themselves leak in known ways. The best documented case is CVE-2025-48757: a researcher tested 1,645 apps built with Lovable and found missing or weak row level security in 170 of them, about one in ten. In those apps, the public key in the JavaScript was enough to read the database. Veracode's 2025 GenAI Code Security Report tested more than 100 language models on 80 coding tasks and found a security flaw in 45% of cases. An engineer may catch those flaws in review. A sales manager has no review. The vibe coding security risks guide lists the eight holes these apps most often ship with.

Why blocking the tools fails

  • The builders run in a browser. Blocking a domain on the office network does nothing on a phone, a home connection or a personal laptop.
  • The app is hosted elsewhere. Once published, it sits on the platform's infrastructure. Your firewall never sees its traffic.
  • A ban moves the work to personal accounts. That is the worst outcome: the app still exists, with company data, and now it appears in none of your records.
  • The benefit is real. The people who build these tools are often the ones who know the process best. A company that forbids them loses the tools and keeps the risk.

The workable position is the one IT took on cloud storage ten years ago: name the allowed tools, make the approved path easier than the hidden one, and keep an inventory. Our vibe coding policy template gives the rules to copy, and the enterprise vibe coding guide covers the review gate. The rest of this page is the inventory.

How to find the apps your people deployed

No single source lists everything. Work through these five in order: the first three are records you already own, and each takes under an hour.

  1. 1

    List the projects in your hosting accounts

    Open the company's Vercel and Netlify teams and the Supabase organization, and export the project list of each. Every project has a production address. Compare the list with what IT knows about. Anything created by someone outside engineering deserves a second look. Do the same for the workspace of each AI builder you pay for.

  2. 2

    Review the apps your people signed in to

    Most builders and hosts offer "Sign in with Google" or Microsoft. Each sign-in with a work account leaves a grant you can read. In Google Workspace, the Admin console lists third-party apps under Security, Access and data control, API controls, with the number of users for each. In Microsoft Entra, the list is under Enterprise apps, All applications, and each app's Permissions tab shows what users consented to. Look for app builders, hosts and backend services, and note who uses them.

  3. 3

    Read the certificate logs for your domains

    Every publicly trusted certificate is recorded in Certificate Transparency logs, which are append-only and open to anyone. An app published on quotes.yourcompany.com needed a certificate, so its name is in the logs. Query them for your own domain and compare the names with your DNS zone. The command below does it.

  4. 4

    Walk through your DNS zone

    Export the records of each company domain. Records that point at a hosting platform are apps or were apps. A record whose target no longer exists is a separate problem: someone else may be able to claim it and serve content under your name. Remove those.

  5. 5

    Ask, with an amnesty

    An app built on a personal account and published under the platform's own domain is in none of the records above. Card statements and expense claims catch the paid ones. For the rest, ask: one form, one month, no blame, and a promise to help rather than delete. People declare tools when declaring them is safe.

bash
curl -s "https://crt.sh/?q=%25.yourcompany.com&output=json" \
  | grep -oE '"name_value":"[^"]+"' \
  | sed 's/"name_value":"//; s/"$//; s/\\n/\n/g' \
  | sort -u
List every host name that got a certificate under your own domain. Replace yourcompany.com.

Write the result in one sheet: address, platform, backend, what data it holds, who built it, who owns it today. An app with no owner is the first finding of the inventory.

What to test on each app you find

An inventory tells you what exists. It does not tell you what a stranger can do to it. For each app, answer five questions from outside, starting with the apps that hold customer or employee data.

QuestionHow to check on your own appDeeper guide
What can a visitor with no account read?On Supabase, list the tables without row level security with the query below. On Firebase, read the rules of each database and bucketSupabase RLS, Firebase security rules
Which keys does the JavaScript carry?Open the app, view the loaded scripts and search them for sk_, service_role and -----BEGINAPI key exposed in frontend
Who can sign up?Try to create an account with a personal email address. An internal tool that accepts it is open to the worldSupabase Auth
Can user B read user A's records?Create two accounts, note a record id from the first, request it as the secondIDOR
Who can call the AI route?Call the route that forwards prompts to a model with no session cookie. If it answers, anyone spends your creditsUnprotected AI endpoint
sql
select schemaname, tablename
from pg_tables
where schemaname = 'public'
  and rowsecurity = false;
Run in the SQL editor of each Supabase project: tables in the public schema with row level security off.

Supabase's documentation is plain on this point: "Enable RLS on every table in an exposed schema." A table returned by that query is readable by anyone who holds the public key, and the public key is in the app's JavaScript by design.

Rank the apps, then decide: adopt, fix or retire

Rank by what each app exposes. An app with customer data and a table open to visitors comes before a calculator with no database. Then take one of three decisions per app.

  • Adopt. The app is useful and passes the five questions. Give it a named owner, move it to a company account, and put it in the inventory.
  • Fix. The app is useful and fails a question. The author's coding agent can usually close the hole in one session. Test again before calling it closed.
  • Retire. Nobody uses it or nobody owns it. Export the data that must be kept, delete the project and its database, revoke its keys and remove its DNS record.

Repeat the inventory on a schedule. Apps are published every week, and a list from last quarter is already wrong.

What Keybrum does for shadow AI, and what it does not

Keybrum is the automated attack test we build for apps made with AI tools. Its Team plan is made for the person who has to answer "which apps did our people publish, and what can a stranger do to them?". You connect your Vercel, Netlify and Supabase accounts and it finds every app deployed there, including the ones nobody declared. Apps are ranked by what they expose, and each gets an owner.

Each app whose ownership is proven is then attacked from outside, the way a stranger would. In plain words: reading Supabase tables, storage buckets and database functions with the public key, reading Firestore collections and Firebase Storage, looking for provider keys in the JavaScript and checking each one with a harmless call, reading one account's records from a second account, calling AI routes with no session, and requesting .env files, .git folders and source maps. Every finding comes with the request that worked. The author's coding agent reads the fix through Keybrum's MCP server, and the finding closes only when the same attack fails.

The Team plan costs €99 a month excluding VAT for 25 apps and 10 members. It attacks every deploy, attacks again every week or month, puts your brand on the reports and gives each app a public trust page that shows what was tested and when. Above 25 apps there is no listed plan: contact us.

Questions

What is shadow AI in simple terms?

Shadow AI is any AI tool used at work without IT approving it or knowing about it. It ranges from a personal chatbot account used for work documents to a full app built with an AI builder and published on the internet. The second kind is the harder one to see, because the app lives outside the company network.

What is the difference between shadow AI and shadow IT?

Shadow IT is any technology adopted without IT: a file sharing account, a project tool. Shadow AI is the part that involves AI tools. The difference that matters today is that AI builders let non-engineers create new software, with its own database and users, where shadow IT mostly meant signing up for existing software.

What are examples of shadow AI?

A recruiter pasting CVs into a personal chatbot account. A finance analyst using an AI browser extension that reads every page. A sales manager who built a quote tool with Lovable and loaded the customer list into it. A developer who gave a coding agent a production database password to move faster.

How do you detect shadow AI?

For chatbot use, network and browser controls show which AI services are reached. For published apps, use three records you already own: the project lists of your hosting accounts, the third-party apps your people signed in to with a work account, and the certificate logs for your domains. An amnesty finds what is on personal accounts.

Should a company ban AI app builders?

A ban rarely works, because the builders run in any browser and the apps are hosted outside your network. It tends to move the work to personal accounts, where you see nothing. A short policy with allowed tools, a rule on data, and a test before publishing gives better results. See the vibe coding policy template.

How much does it cost to test every app employees built?

With Keybrum, the Team plan is €99 a month excluding VAT for up to 25 apps, with discovery of the apps deployed on Vercel, Netlify or Supabase. A single app can be tested once with a one-time report at €49. Only the check at /check, the seven tools and the sample report cost nothing.

Sources

  1. 1.What is shadow AI?, IBM
  2. 2.2024 Work Trend Index: AI at work is here. Now comes the hard part, Microsoft and LinkedIn
  3. 3.IBM Report: 13% of organizations reported breaches of AI models or applications (Cost of a Data Breach Report 2025), IBM Newsroom
  4. 4.CVE-2025-48757, NIST National Vulnerability Database
  5. 5.Statement on CVE-2025-48757, Matt Palmer
  6. 6.AI-generated code poses major security risks in nearly half of all development tasks (2025 GenAI Code Security Report), Veracode
  7. 7.Control which apps access Google Workspace data, Google Workspace Admin Help
  8. 8.Review permissions granted to enterprise applications, Microsoft Learn
  9. 9.How Certificate Transparency works, Certificate Transparency
  10. 10.Row Level Security, Supabase Docs

Your app is public. So are its mistakes.

Check what any attacker can already see in it, free, in ten seconds. The full attack, with proof and fixes: €49 once, or €29 a month on a plan.

or continue with Google or GitHub