[email protected]

Basics · Security & Pentesting

Vibe Coding Risks: Vulnerabilities in AI-Generated Code and How to Check for Them

Editorial note: This guide was written to the best of our knowledge at the time of publication. Versions, commands, features, prices and licensing terms may change through vendor updates; the current information from the respective provider applies. Please verify critical steps independently before production use. All trademarks, product names and company names mentioned belong to their respective owners. As a rule, WZ-IT is not a partner of the vendors and projects mentioned and is not officially affiliated with them; existing partnerships are stated explicitly. This text does not replace individual professional, legal or tax advice. If any information is outdated or incorrect, please write to [email protected].

Have your vibe-coded application reviewed - We check sign-in, access rights, secrets and dependencies of your application in the code and on a test instance. Confirmed findings come with proof. Every finding carries one of four states: confirmed, refuted, suspicion without confirmation or not tested. Book an intro call · All review levels at a glance

Vibe coding means creating software from a description in natural language without reading the code in detail. Whether the result also handles access rights, secrets and inputs properly only shows in a review. This article summarizes what studies and authorities document, which vulnerability classes the sources name and how to check an application built by prompt.

All figures come from the original sources, with sample size and year. Studies measure models on test tasks. They do not say how secure a specific application is.

Table of contents

Vibe coding risks at a glance

Question Answer
What is vibe coding? Creating software from a description and judging it on the running system. Andrej Karpathy described and named the way of working in February 2025 (post on X)
How often does AI code contain vulnerabilities? Depends on the study and method: NYU 2021 about 40 % of the programs; Georgetown CSET 2024 about 48 % of the C code snippets; BaxBench 2025: best model 62 % correct solutions, exploits executable on around half of each model's correct programs. The values are not comparable and apply to test tasks, not to a specific application
Which vulnerability classes should be checked? Missing access check per record, secrets in code, missing input validation, overly broad database and CORS rules, invented dependencies, missing tests. No source documents a ranking by frequency
Who is responsible? Not answerable in general terms. BSI and ANSSI: AI assistants are no substitute for experienced developers (publication). The classification in a specific case is a matter for legal counsel
What helps? Access test with two accounts, scan for secrets and dependencies, configuration check. With real data, additionally audit and pentest (our assessment)

What is vibe coding and which tools are meant?

In February 2025, Andrej Karpathy described in a post on X "a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists". According to the Fraunhofer IESE (article in German), he coined the term with it.

In the narrow sense, vibe coding means that only the running system is judged and the code is neither understood nor reviewed. The Fraunhofer IESE considers this way of working unsuitable for developing professional software. When experts supervise the generated code, it speaks of AI-assisted software engineering.

This article uses the term broadly: applications built mostly by prompt and put into operation without a complete code review. The following grouping of tools is our own classification.

Type of tool Examples Typical result
Platforms that create a running application from a description Lovable, Bolt, v0, Replit, Base44 Frontend, backend connection, often with hosting on the platform
Assistants and agents in the editor or terminal Cursor, Claude Code, GitHub Copilot Changes directly in the repository

The joint publication by BSI and ANSSI (last updated September 2024) covers AI coding assistants in professional software development. No-code and low-code applications are explicitly out of its scope.

What studies show

The sources measure different things: model output on test tasks, the behavior of developers and findings on real applications. The tables list sample, finding and limits for each source (status: November 2026).

Measurements of AI-generated code

Source and year Sample and method Finding Limits
NYU: Asleep at the Keyboard, 2021 89 scenarios on weaknesses, including from the CWE Top 25, 1,689 programs generated by GitHub Copilot About 40 % of the programs were vulnerable Copilot as of 2021, scenarios designed to target weaknesses
Stanford: Do Users Write More Insecure Code with AI Assistants?, ACM CCS 2023 User study with 47 participants, five security-related tasks in Python, JavaScript and C, assistant based on OpenAI codex-davinci-002 With an assistant, participants wrote insecure solutions more often in four of five tasks and were more likely to believe their code was secure Small sample, older model generation
Georgetown CSET, November 2024 Five models (GPT-4, GPT-3.5-turbo, Code Llama 7B Instruct, WizardCoder 7B, Mistral 7B Instruct), 67 prompts for C code (Code Llama delivered only 62 usable snippets), checked with the verification tool ESBMC About 48 % of the code snippets compiled and contained a flagged bug, about 30 % passed verification C only. The authors call the result a rough upper bound and limited in scope
BaxBench (ETH Zurich and others): Can LLMs Generate Correct and Secure Backends?, ICML 2025 392 tasks (28 scenarios in 14 backend frameworks and 6 programming languages), functional tests and executed exploits against the generated applications The best model (OpenAI o1) reached 62 % on correctness. On around half of the correct programs of each model on average, security exploits could be executed successfully Benchmark tasks, model state of 2025
Spracklen et al., USENIX Security 2025, arXiv 2024 (version of March 2025) 16 models, 576,000 code samples in Python and JavaScript, 2.23 million generated package mentions 440,445 (19.7 %) classed as invented, including 205,474 unique names. Average at least 5.2 % for commercial and 21.7 % for open-source models Measures invented names, not actual attacks

A documented finding on a platform

Source and year Sample and method Finding Limits
NVD: CVE-2025-48757, published 30 May 2025 CVE entry According to the entry, an insufficient row-level security policy in Lovable through 15 April 2025 allowed unauthenticated attackers to read or write arbitrary database tables of generated sites. CVSS 3.1 base score 9.3 (rated by MITRE) The entry notes that the supplier disputes the finding. Timing and scope are not independently confirmed. A single case that says nothing about frequency

What authorities and institutes state

  • BSI and ANSSI (last updated September 2024): Mild and severe security flaws are commonly present in AI-provided code snippets. Quality varies widely by programming language and task. Generated code should be checked and reproduced by developers. The productivity gain of development must be compensated by more capacity in quality assurance (publication).
  • Fraunhofer IESE (27 November 2025): The blog article (in German) lists eight risks, including missing security checks, hard-to-maintain code, missing tests and invented dependencies. It contains no measurements of its own.

What can be concluded

  • Measurements from 2021 to 2025 find insecure output in a considerable share of test tasks (about 40 % at NYU, about 48 % at CSET; at BaxBench, exploits could be executed on around half of the correct programs). The values measure different things (tasks, languages, models, definition of failure) and do not add up to a common failure rate.
  • Running code is not secure code. In BaxBench, security exploits could be executed on around half of the functionally correct programs of each model on average.
  • Self-assessment can mislead. In the Stanford study (47 participants, five tasks), participants with an assistant were more likely to believe their code was secure, although they wrote insecure solutions more often in four of five tasks.
  • What cannot be concluded is how common insecure applications are in practice. None of the studies measures applications of a specific platform.

Vulnerability classes in AI-generated code

The selection and mapping of the classes is our own compilation. The classification follows the OWASP Top 10:2025 and the CWE entries by MITRE. None of the cited sources documents a ranking by frequency in real vibe-coding applications. The examples are patterns, not individual cases.

Risk Example Classification Check
Missing access check per record A route returns invoice 123 to any signed-in user because it checks only the sign-in, not ownership A01:2025 Broken Access Control, CWE-862 Missing Authorization, CWE-639 Two test accounts (different role, different tenant). Call every route with foreign IDs and without sign-in
Overly broad database rules Table without row level security. According to the Supabase documentation it is readable and writable by any role with a grant on it in an exposed schema A01:2025 Broken Access Control, CWE-863 Incorrect Authorization (as in the NVD entry) Check RLS and grants per table in exposed schemas, read the policies, test access with the public key for reading and writing
Secrets in code Secret or privileged key (secret key, service_role, payment or service key) in the frontend bundle or in the git history. An explicitly public client key is not a vulnerability by itself CWE-798 Use of Hard-coded Credentials, A07:2025 Authentication Failures Scan the repository including history, search the frontend bundle, rotate found keys instead of only deleting them
Outdated or invented dependencies (slopsquatting) A model recommends a package that does not exist. An attacker registers the name A03:2025 Software Supply Chain Failures, OWASP LLM09:2025 (insecure or non-existent libraries) Check every dependency against the registry (exists, age, repository), pin a lockfile, vulnerability scan, software bill of materials (SBOM)
Missing input and output validation A form value ends up unchecked in a query or page A05:2025 Injection, CWE-20, CWE-79, CWE-89, CWE-117 (A09:2025) Server-side validation against a schema, parameterized queries, output encoding, tests with malicious values
Overly broad CORS and configuration rules On a non-public API, CORS allows any origin or reflects the request origin together with cookies. A wildcard origin without credentials can be acceptable for deliberately public resources. Debug mode and open admin routes in production A02:2025 Security Misconfiguration, CWE-942, WHATWG Fetch: CORS and credentials Check response headers of the production environment (origin rule, credentials), compare development and production configuration
Weak cryptography Passwords stored with MD5 or a single SHA-256 round. BSI and ANSSI name this as an example of outdated methods that assistants still suggest A04:2025 Cryptographic Failures Check password storage and algorithms in the code, move to current methods
Missing tests and logging No tests for permissions. A later change by prompt silently removes a protection No category of its own. Closest are A09:2025 Security Logging and Alerting Failures (CWE-778 Insufficient Logging) and A06:2025 Insecure Design (our mapping) Automate permission tests per role and integrate them into delivery. BSI and ANSSI recommend automated function tests

On repeatability: according to the Stanford study, errors can multiply when earlier outputs of the assistant are used as new prompts. The Fraunhofer IESE names missing or incomplete tests as one of eight risks.

What changes organizationally

Responsibility

BSI and ANSSI state that AI assistants are no substitute for experienced developers and that unrestrained use can have severe security implications. The NVD entry for CVE-2025-48757 notes that the platform supplier disputes the finding because each customer accepts responsibility for protecting the data of their application. That is the supplier's position, not a legal ruling.

According to the European Commission, the EU Product Liability Directive 2024/2853 clarifies that software counts as a product, including applications and AI systems. The Commission names the manufacturer of the product as the primary liable party. The directive does not cover cases in which the injured party is a company. It applies to products placed on the market after 8 December 2026, and member states must transpose it by 9 December 2026. The directive excludes free and open-source software that is developed or supplied outside the course of a commercial activity (Art. 2(2), Directive 2024/2853). Whether an economic operator is liable for a specific application depends on its role in the individual case. That is a matter for legal counsel.

Code ownership

In its report of 29 January 2025, the US Copyright Office concludes that copyright does not extend to purely AI-generated material and that, based on current technology, prompts alone do not provide sufficient control over the expression. In Germany, under section 2 (2) of the Copyright Act (UrhG) (German text), only personal intellectual creations are works. For computer programs, section 69a (3) UrhG (German text) applies: they are protected if they are individual works, that is, the result of the author's own intellectual creation. Whether source code created by prompt is protected has to be assessed case by case. The sources address only eligibility for protection. Authorship, ownership, contractual rights of use and license terms have to be assessed separately.

Our assessment for practice: the complete code sits in your own repository, the terms of use of the tool have been read, and operation does not depend on an account with the platform.

Traceability

The Fraunhofer IESE describes declining understanding of the system as a risk: without insight into the code, it is unclear without extensive testing whether the software has the intended properties. BSI and ANSSI name as possible measures flagging AI-generated code blocks, documenting the AI tools used and creating a software bill of materials (SBOM).

Maintenance

The Fraunhofer IESE describes AI-generated code as often unstructured, redundant or hard to maintain. BSI and ANSSI point out that developers should be able to continue operations without AI assistance, for example in an incident. Our assessment: over time the application needs dependency updates, review of new changes and someone who fixes findings. For this there is continuous AI code review and takeover into operations.

How to check an application built by prompt

Steps for checking it yourself

The steps are our own compilation, not a requirement of a standard. They can be taken without external help. For more depth, the OWASP Application Security Verification Standard 5.0.0 (published 30 May 2025) serves as a checklist catalog.

  1. Record the inventory: repository, hosting, database, third-party services, keys and who has administrator access. Check whether the code can be exported completely.
  2. Search for secrets: scan the repository including history with a secret scanner such as Gitleaks and look at the delivered frontend bundle in the browser. Rotate any keys found. A key meant for the browser (public Supabase key) is something different from a secret key, which must never be in the browser.
  3. Check dependencies: compare every dependency against the registry (exists, plausible, maintained). Search for known vulnerabilities with a tool such as the OSV-Scanner.
  4. Test access rights: create two accounts per role and tenant. With account A, call the IDs and routes of account B, for reading and writing. Call routes without signing in. With Supabase, check RLS and grants per table and read the policies.
  5. Test inputs: fill forms and API fields with malicious values (special characters, script tags, query fragments). Validation must happen server-side.
  6. Check configuration: CORS, debug modes, error pages, response headers, open admin or preview URLs and storage buckets of the production environment.
  7. Secure tests and operations: automate permission tests per role. Try backup and restore, set up logging and alerting, clarify who responds.
  8. Repeat: check again after every major change made by prompt, because any change can alter a protection.

When an external review makes sense

Our assessment, not a requirement of a standard:

  • The application processes personal data, payments or data of several tenants.
  • It is reachable from the internet and has accounts with different roles.
  • Nobody on the team can explain how sign-in and access checks work.
  • A customer or partner asks for an independent review report.
  • The application goes live, is handed over to another team or is taken over into operations.
  • There have been major changes by prompt since the last review.

Scan, audit and pentest compared

Method Question Approach Result
Vulnerability scan Which known vulnerabilities are detectable? A tool identifies systems and associated vulnerabilities (NIST SP 800-115) List of suspected issues, human assessment needed
Audit (code review) Is the application ready for operation? Review of code, data access, configuration and operations against a checklist (term as we use it) Prioritized list of measures
Pentest What can actually be exploited? Mimicking real-world attacks to circumvent security features (NIST SP 800-115) Findings with proof, severity and recommendation

More on the distinction in the knowledge article AI code audit, code review or penetration test.

Common misconceptions

Misconception Assessment Source
If the application runs and looks good, it is secure Running does not mean secure. In BaxBench (2025) the best model produced only 62 % correct backends, and security exploits could be executed on around half of the correct programs of each model BaxBench
With an AI assistant I write more secure code In the Stanford study (47 participants, five tasks), participants with an assistant wrote insecure solutions more often in four of five tasks and were more likely to believe they were secure Perry et al., CCS 2023
Sign-in protects the data Sign-in is not authorization. Without a check per record, any signed-in account can retrieve foreign data. For directly exposed tables without a suitable RLS policy, access can also be possible without signing in, depending on grants, as in the entry for CVE-2025-48757 OWASP A01:2025, Supabase, NVD
A key in frontend code is always a vulnerability It depends on the key. A public Supabase key is meant for the browser but does not protect any data by itself: access is limited by grants and RLS policies on the tables. A secret key must never be in the browser Supabase RLS, Supabase API keys
Slopsquatting is pure theory Invented package names are measured: 19.7 % of package mentions in a study with 16 models. The study does not document actual attacks, however. BSI and ANSSI list the attack path as a risk Spracklen et al., BSI and ANSSI
A review by an AI is enough for AI-generated code Scanners and AI questioning reduce risk. BSI and ANSSI additionally recommend review by developers BSI and ANSSI
A scan replaces a pentest A scan identifies known vulnerabilities. A pentest mimics attacks, and most pentests also look for combinations of vulnerabilities NIST SP 800-115
Whoever writes the prompt automatically owns the copyright In the view of the US Copyright Office, prompts alone do not provide sufficient control. In Germany the personal intellectual creation counts US Copyright Office, UrhG section 2, UrhG section 69a

Status and methodology

Status: 19 November 2026. The figures come from the original sources: studies (arXiv, publications of universities), publications by BSI, ANSSI and Fraunhofer IESE, standards texts and database entries (OWASP, MITRE, NVD, NIST). The attribution of the term vibe coding to Karpathy rests on the account of the Fraunhofer IESE (a secondary source) and on his post on X of 2 February 2025. For the figures, reports by providers of security tools and security testing, media reports and secondary quotations are not used as evidence.

Limits: the studies measure models and tools of their time, and new versions shift the values. Figures from different studies are not comparable (different tasks, languages, models, definitions of failure). The study by Perry et al. uses a small sample, the NYU study an early Copilot version, BaxBench models of 2025. The selection and mapping of vulnerability classes, the self-check steps and the thresholds for an external review are our own compilation. Statements marked "Our assessment" are not evidence from sources. The section "Our approach at WZ-IT" describes our own offering and is not a statement from sources. Legal notes are not legal advice.

Our approach at WZ-IT

  1. Handover: you hand over source code, a test instance and test accounts per role and tenant. Ownership of the application is verified.
  2. Inventory: we record the entry points and routes of the application. The runtime view of the routes is currently available for Express (Node.js).
  3. Audit: the Production Readiness Audit checks sign-in, data access, security, tests, deployment and operations and delivers a prioritized go-live list.
  4. Pentest: the automated whitebox pentest for web applications and APIs attacks the test instance, never the production system. Confirmed findings come with proof. Every finding carries one of four states: confirmed, refuted, suspicion without confirmation or not tested. The team reviews severe confirmed findings by hand. The report includes a coverage table.
  5. Fix and operate: on request we fix the findings and take over hosting and operations (takeover, operations). The continuous AI code review accompanies changes made by prompt.

Prices, net, as of November 2026:

Service Price
Pentest, one-off €1,000 (small), €1,500 (medium), €2,000 (large), net, fixed price by size of the application
Pentest subscription, quarterly €2,500 (small), €3,750 (medium), €5,000 (large), net per year
Pentest subscription, monthly €4,000 (small), €6,000 (medium), €8,000 (large), net per year
Production Readiness Audit from €1,490
Combined package audit plus pentest from €1,990 for a small application
AI code review, continuous from €990 per month

Further guides

Built by prompt and real data involved? We review code, data access and operations and attack a test instance. Book an intro call · Vibe coding security audit

Sources

Enquiry

Have your application assessed

Describe the application, roles and baseline. We assess whether an audit, a pentest or both fits, and state the price.

  • You speak directly with Timo and Robin, who are technically responsible
  • A professional assessment of your situation
  • Concrete next steps for infrastructure, software or AI

The conversation is free and non-binding.

I got to know Timo and Robin as dedicated and professional partners who make a lot of things possible for their clients that initially seem impossible. I felt very well advised. The services offered have thoroughly convinced me. (translated)
Andrea PawlowskiCommunications Manager, Golem.de, Germany
Timo and Robin, founders of WZ-IT

Have your application assessed

I got to know Timo and Robin as dedicated and professional partners who make a lot of things possible for their clients that initially seem impossible. I felt very well advised. The services offered have thoroughly convinced me. (translated)
Andrea PawlowskiCommunications Manager, Golem.de, Germany

Frequently Asked Questions

Answers to the most important questions

Customers from mid-sized business, industry and the public sector, in Germany and internationally

  • Stadtwerke Brühl
  • DGHO e.V.
  • solarwerkstatt berlin
  • Krüger Dirndl
  • ABCO Water Systems
  • Golem.de
  • EVADXB
  • nextGYM
  • AInergy
  • ml&s
  • Odiseo Solutions
  • Annota
  • ARGE
  • SweetConnect GmbH
  • Aphy AG
  • CORGOS
  • Rekorder
  • SolidProof
  • Yonju
  • Keymate
  • Paritel
  • Mr. Clipart
  • Millenium
  • Negosh
  • Führerscheinmacher
  • BTI
  • Deutsche Gesellschaft für Datenschutz
Read client reviews