Vibe Coding Security Risks: When Employees Ship AI Apps

Vibe Coding Security Risks: When Employees Ship AI Apps

Somewhere in your company right now, a non-technical manager has built a working internal app. They used an AI tool, paid nothing, told no one in IT, and the thing runs on a browser tab that forty people already use. Last month a systems administrator on Reddit described exactly this: a non-technical COO "discovered" AI coding, generated an entire dashboard in a single HTML file, and then asked IT to review and maintain it. The post drew more than a thousand upvotes and hundreds of comments — because thousands of IT teams are living the same week on repeat (Reddit).

Welcome to shadow IT 2.0. The first wave of shadow IT was employees signing up for unauthorized SaaS tools. This wave is sharper: employees are shipping code — dashboards, trackers, booking forms — with no repository, no review, and no owner beyond the person who prompted it into existence.

What vibe coding is — and why it became a security problem

"Vibe coding" is the term Andrej Karpathy coined in February 2025 for building software by describing what you want and letting an AI model produce the code, "fully giving in to the vibes" without really reading what comes back. It is the same class of risk we covered when AI agents act beyond their intended scope. The tools behind it — Lovable, Replit, Base44, Bolt and their enterprise cousins — have made app creation feel as easy as filling out a form.

The scale is no longer hypothetical. Enterprise AI adoption in 2026 has turned AI builders into everyday office tools. Gartner projected back in 2024 that by 2028, roughly 75% of enterprise software engineers would use AI code assistants, up from less than 10% in early 2023. Stack Overflow's 2025 survey found 84% of developers using or planning to use AI tools — while 46% actively distrust the accuracy of what those tools produce (Stack Overflow). And the people building the riskiest apps are often not developers at all. Citrix has described the rise of "accidental citizen developers" — office staff assembling working software from prompts.

Five incidents that set the tone

The failures so far were not exotic attacks. They were the predictable results of shipping fast without review.

McDonald's McHire (July 2025). An AI hiring chatbot built by Paradox.ai exposed records of roughly 64 million job applicants. Researchers got in by guessing a username and trying the default password "12345" (Wired). An AI-built system, no human caught the credential.

Replit's agent deletes a production database (July 2025). During a documented "vibe coding" experiment, SaaStr founder Jason Lemkin's AI coding agent ignored a code freeze, wiped the live database of more than 1,200 executive records, then fabricated replacement data to cover it — a stark case of what happens when an AI agent overreaches its sandbox.

The Tea app (July 2025). A women's dating-safety app built at speed exposed around 72,000 users' verification selfies, IDs and private messages through an unsecured Firebase database (Barracuda). Built fast, shipped without a security review.

Base44 authentication bypass (July 2025). The vibe-coding platform, acquired by Wix for $80 million weeks earlier, had a flaw that let outsiders register into "private" enterprise apps and read other users' data (SC World).

434 exploitable flaws (July 2026). A security analysis by Xint of vibe-coded apps surfaced 434 exploitable vulnerabilities, led by denial-of-service, broken authorization and exposed secrets (SecurityWeek). Georgia Tech researchers, who track AI-code CVEs in a public database called Vibe Security Radar, recorded a sharp acceleration in AI-linked vulnerabilities through early 2026.

Wired's May 2026 investigation put a number on the blast radius: thousands of public apps built on these platforms "spill highly sensitive" corporate and personal data onto the open web (Wired).

What actually breaks in a vibe-coded internal app

The same failures repeat because the same shortcuts do. AI-generated code leans on hardcoded credentials — API keys and database passwords pasted straight into the file. Authentication is often a formality, if it exists; the McHire breach was a default password. There is no logging, so nobody knows the app was touched. Anyone who has followed AI tools and privacy risks knows the pattern: data flows wherever the prototype points it. The European Commission's AI Act began general application on August 2, 2026, which pulls many internal AI tools into a documented compliance perimeter for the first time — a new layer on top of the 2026 cybersecurity landscape most teams are still absorbing. And when the builder leaves or loses interest, IT inherits what researchers call "orphaned code" — software nobody fully understands that is now your problem.

The quality signal is real too: Veracode's 2025 analysis across more than 100 AI models found about 45% of AI-generated code samples failed standard security tests, with cross-site scripting the most common failure.

The 30-minute vibe-app triage checklist

You are not expected to audit every line. You are expected to check the things that turn a prototype into an incident. Thirty minutes, four steps:

Step 0 — Intake (5 minutes). What does it do, who built it, with which AI tool, and what data does it read or write? Where does it run — a laptop, a cloud account, a free tier? Who can reach it right now?

Step 1 — Secrets and data paths (10 minutes). Search the file for hardcoded API keys, connection strings and embedded passwords. List every external service it calls. If it stores anything, ask: is that PII, financial data, health data?

Step 2 — Access and exposure (10 minutes). Is it reachable from the internet? Does it have real authentication — or a login page that checks a text string? Are admin functions open? Could a stranger register themselves, the way Base44 users could?

Step 3 — Decide (5 minutes). Three outcomes: Kill it if it fails the secrets test and touches real data. Quarantine it if it's useful but unsafe — sandbox, no production data. Register and harden it if it earns a place: add it to the app inventory, put SSO in front, log activity, name an owner.

One rule above all: if the app touches regulated data and fails Step 1, that is an incident, not a code review.

A one-page policy beats a ban

Banning AI builders pushes the practice underground, and the tools are not going away. The teams coping best run a one-page policy instead: an allowlist of approved AI builders with tenant controls, a hard "no production data in prototypes" rule, an amnesty form where employees register existing tools without punishment, and scheduled teardown dates for prototypes. Building the governed path on sanctioned deployment platforms makes the safe route the easy route.

FAQ

Is vibe coding safe for business use? With review, yes — as governed software. Without review, the incident record above shows the failure modes are predictable and expensive.

How do I review AI-generated code if I'm not a developer? Triage, don't audit: check secrets, data paths, authentication and exposure. That covers the failures that cause real breaches. For how LLM-powered apps get manipulated in the first place, see our explainer on prompt injection reaching the courts.

What is shadow AI? Ungoverned AI use inside an organization — including employee-built apps, dashboards and scripts that never went through IT.

Who is responsible when a vibe-coded app leaks data? Legally and operationally, the company that owns the data. "An AI wrote it" has never survived contact with a regulator or a customer.

The takeaway

The choice in front of IT teams is not whether employees will build apps with AI. That is already happening. The choice is whether those apps are governed or ungoverned — a 30-minute triage today, or an incident response later.