Is AI-Written Code Secure? A 10-Point Acceptance Checklist

Research shows about 45% of AI-generated code samples fail OWASP Top 10 tests, and AI-linked vulnerabilities rose fast in 2026. The common flaws and a 10-point acceptance checklist.

The Situation: Fast Is Not the Same as Secure

Writing code with AI is now normal. Whether you built an internal system with a tool yourself or your vendor uses AI to deliver faster, the system in your hands very likely contains AI-generated code. That is not a problem in itself. The problem is whether checking kept pace with the speed.

A research note published by the Cloud Security Alliance in April 2026 pulls together figures every buyer should know:

Real vulnerabilities are rising too. Georgia Tech's Vibe Security Radar scanned more than 43,000 security advisories and confirmed 74 cases directly linked to AI-generated code, 14 critical and 25 high. March 2026 alone had 35, more than all of 2025. TechNews reports the 74 as of March 20, 2026, and the researchers estimate the true number may be 5 to 10 times higher because the tool only recognizes code that carries clear AI signatures.

The Three Most Common Flaws

The CSA note identifies three dominant failure patterns. They share one trait: the program runs and the screens look fine, and the flaw appears only when someone deliberately probes for it.

FlawIn plain languageLikely consequence
Authorization logic errorsAfter logging in, changing an ID in the URL shows someone else's orderCustomers can see each other's data
Missing access controlsThe database or admin interface has no permissions, so anyone with the address can read and writeThe whole database leaks or is altered
Hardcoded credentialsConnection passwords and payment keys written into the code, sometimes in the front endThird-party services are abused, costs accrue or data leaves

Georgia Tech's tracking adds three technical categories: command injection, authentication bypass and server-side request forgery. Non-technical decision makers need not memorize the terms. Remember one thing: AI is good at making a feature work, and it will not ask who should be prevented from using it.

Acceptance Checklist: 10 Questions Anyone Can Ask

For each item, ask the vendor or your engineers to demonstrate or provide evidence on the spot. Anything they cannot answer needs fixing before acceptance.

Permissions and data

  1. Logged in as user A, can you see user B's data? Have them change the ID in front of you.
  2. Can someone who is not logged in open the admin URL or data interface directly?
  3. Are user and administrator functions separated? Can a normal account delete, export or change prices?
  4. Does the database accept connections only from the system itself? It should not be reachable from outside.

Keys and configuration

  1. Are there passwords or keys in the code? Ask for the output of a secret-scanning tool.
  2. Were test accounts, default passwords and debug features disabled before launch?

Dependencies

  1. Which third-party packages are used? Ask for the list and whether a vulnerability scan was run.
  2. Does the list include obscure packages with very few downloads or a recent registration date? That is the typical mark of a hallucinated package.

Process

  1. Did a person review the AI-generated code? Login, payment, permissions and external interfaces should be reviewed by someone with security experience.
  2. Is there an automated security scan before release? On every update, or only once?

The Georgia Tech researchers give practical advice: review AI-generated code the way you would review a junior developer's pull request. It may be good. It still should not ship just because it looks good.

Questions to Ask a Vendor

The CSA recommends updating vendor questionnaires to ask whether applications were AI-generated. The aim is not to reject AI but to confirm a matching review process exists. Before signing, ask:

For a full evaluation method, see our software vendor checklist. For how AI changes development cost and quality, see AI-assisted development costs.

What If the System Is Already Live?

If your system was built quickly with AI tools and never went through the checks above, work in this order:

  1. Scan for secrets first: fastest to do and most likely to be exploited immediately. Rotate anything exposed.
  2. Check database access settings: confirm that not everyone can read and write.
  3. Run a round of permission tests: try accounts of different roles against each other.
  4. Inventory and update packages.

If the system already holds real customer data, an independent assessment is worth the cost. When the problems are extensive and the original vendor cannot fix them, our guide to taking over an outsourced system helps you choose between patching and rebuilding. For everyday protection, see our SME cybersecurity checklist.

Next Step

AI made a working system cheaper to produce. The cost of making it secure did not disappear. It moved from writing to reviewing. Asking these 10 questions at acceptance costs far less than handling one data breach after launch.

EFFECT uses AI assistance in development while treating permission design, key management and pre-launch checks as fixed process steps, and we also review existing systems. Bring your system or acceptance documents to a free 30-minute consultation (NDA protected) and we will help you see what to fix first.

Sources

FAQ

Is AI-written code less secure than human-written code?

Current research does not support that conclusion directly. What is established is that in Veracode's testing about 45% of AI-generated samples failed OWASP Top 10 security tests, and the rate did not improve across model generations. The practical difference is speed: AI produces code faster, and if review and testing are not strengthened to match, more unchecked code reaches production.

I cannot read code. How do I accept a system on security?

You do not need to read code. Ask for demonstrations and evidence: two different accounts trying to view each other's data, an attempt to open the admin URL without logging in, secret-scan and vulnerability-scan results, and the list of third-party packages. Anything the vendor cannot answer or refuses to demonstrate needs fixing before acceptance.

What is slopsquatting, and why should I care?

AI sometimes suggests package names that do not exist. Attackers notice which names are commonly invented, register packages under them and plant malicious code, so a developer who installs what the AI suggested is compromised. A study of 576,000 samples found 19.7% of AI-suggested dependencies did not exist. The defense is to confirm a package exists, comes from a trustworthy source and has a reasonable publication history before installing it.

Does a small internal tool we built with AI need this much scrutiny?

It depends on the data it touches. A tool used only inside the company with no personal or financial data carries relatively low risk. Once a system is on the internet, stores customer data or connects to payments and other systems, at minimum run the secret scan, check database access settings and test permissions.

Bring us one workflow that keeps getting stuck

No complete specification required. A 30-minute first call clarifies the problem, data and desired outcome. Project ideas remain confidential.

Book a 30-Minute Call