Vibe coding means building software by describing what you want to an AI and accepting what it writes without closely reading the code. Andrej Karpathy coined the term on 2 February 2025 for throwaway weekend projects. The evidence is mixed: speed gains are uneven, while security and understanding suffer when nobody reviews the output, so the safe version adds a spec, small steps, tests and review.
Vibe coding went from a single post to a dictionary entry in under a year. The name is now used for very different things, from a weekend toy to a professional workflow with coding agents. This page separates them and shows what the research and the incidents say, with the source for every number.
- 1Describesay what you want
- 2Generatethe AI writes code
- 3Runsee what happens
- 4Reviewread and test it
- 5Fixsmall, checked changes
What vibe coding means
Vibe coding is writing software by telling an AI what you want, running what it produces and asking for changes, without reading or fully understanding the code. Andrej Karpathy introduced the phrase in a post on X on 2 February 2025. He described giving in to the vibes and "forget that the code even exists": asking for changes by voice, accepting everything, pasting error messages back with no comment. He said it was fine for throwaway weekend projects.
The word spread fast. Collins Dictionary named it Word of the Year for 2025, defining it as using AI, prompted in natural language, to help write computer code. That broader meaning is why people argue about it: Karpathy's original described the careless end, and the dictionary covers almost any use of AI to write code.
A spectrum, not one thing
Simon Willison drew a useful line in March 2025: keep the term for building software with an LLM without reviewing the code it writes. If you review the output, test it and could explain how it works, that is software development with an AI tool. Kent Beck calls the careful version augmented coding, where you still care about the code, its complexity and its tests. In February 2026 Karpathy wrote that programming through agents was now becoming a default professional workflow with more oversight, and suggested "agentic engineering" for it.
| Mode | Do you read and test the code? | Best for | Where the term comes from |
|---|---|---|---|
| Vibe coding | No | Throwaway prototypes and weekend projects | Karpathy, 2025; Willison's definition |
| Augmented coding | Yes: you care about the code, tests and complexity | Everyday development with AI help | Kent Beck, 2025 |
| Agentic engineering | Yes: you specify, supply tools, verify and iterate | Production software built with coding agents | Willison and Karpathy, 2026 |
This hub uses the spectrum. Learn to vibe code for prototypes, then move toward the reviewed end whenever real users or real data are involved.
What vibe coding is good for
- Prototypes you will show to three people and might throw away.
- Internal tools, scripts and one-off automations where you are the only user.
- Learning what is possible before you invest in a proper build.
- Getting a first version in front of users quickly, then rebuilding what survives.
- Exploring a new framework by seeing working code, as long as you also read it.
What the evidence says
The research is mixed, and it is easy to find a number that supports any view. These are the main findings, with who ran each one. Several come from companies that sell security or AI products, which is marked.
| Finding | Figure | Run by |
|---|---|---|
| Experienced developers took longer with early-2025 AI tools, yet believed they were faster | 19% longer across 246 tasks (16 developers); they expected to be 24% faster and later estimated 20% faster | METR, independent research |
| A later METR update could not size the effect | Pointed the other way, but METR calls it unreliable because developers now refuse to work without AI | METR update, February 2026 |
| Few professionals vibe code | 15.1% of 26,564 respondents said it is part of their professional work | Stack Overflow Developer Survey 2025, self-reported |
| Trust is low | 32.7% trust AI accuracy and 45.7% distrust it; 45.2% say debugging AI code takes longer | Stack Overflow Developer Survey 2025 |
| AI code often fails security tests | 45% of samples failed in 2025; a 56% pass rate in 2026, rising to 62% in an October update | Veracode, a security vendor; tested raw models |
| Vulnerabilities traced to AI-written code | 74 confirmed CVEs by March 2026; a floor, not a total | Georgia Tech Vibe Security Radar, university research |
| Delegating everything weakens learning | 50% vs 67% on a later quiz, with no significant speed-up; asking conceptual questions kept scores up | Anthropic, an AI vendor; small preprint |
| AI invents package names | 19.7% of 2.23 million recommended packages did not exist (older models, 2025 paper) | Academic study, USENIX Security 2025 |
A fair reading: AI tools can speed up some work, they are widely used, and professionals mostly keep a human in the loop. Security results are poor without review, and the biggest risk to beginners is not understanding what they shipped.
Incidents worth knowing
- Replit and SaaStr, July 2025. SaaStr's founder reported that Replit's AI agent deleted his production database during a code freeze that existed only as an instruction. Replit's CEO called it unacceptable and announced automatic separation of development and production databases. See The Register's report.
- Lovable and CVE-2025-48757, 2025. Many Lovable-generated apps used database tables without row-level security, so anyone with the public key could read or change data. Lovable's side is that each customer is responsible for protecting their own app's data, and the record notes that dispute.
- Moltbook, January to February 2026. Wiz found a database key in browser code and no row-level security, exposing about 1.5 million authentication tokens and 35,000 email addresses. It was secured within hours.
- Base44, July 2025. Wiz found two endpoints that let anyone register for a private app knowing only its app ID. This was a platform bug, fixed within a day, with no evidence of exploitation.
- RedAccess, May 2026. The firm claimed about 5,000 publicly reachable vibe-coded apps with little or no authentication. The platforms disputed the framing, and the count is the firm's own claim, but several exposed apps were verified as real.
The pattern is consistent: missing access rules, secrets in browser code and apps that are public by default. The Tea app breach is often blamed on vibe coding, but Tea attributes it to legacy storage and there is no evidence it was vibe coded.
When not to vibe code
- Login, roles and anything that decides who can see what. Broken access control is OWASP's number one risk, and the headline incidents above were access failures.
- Payments and card data. Use a hosted checkout and verify signatures; do not let an AI improvise a payment flow.
- Health, finance, ID documents and other sensitive personal data. India's Digital Personal Data Protection Act, 2023 allows penalties of up to Rs 250 crore for failing to take reasonable security safeguards. This is general information, not legal advice.
- Anything real users depend on, unless you review and test it. Karpathy scoped his original to throwaway projects.
- Code you must maintain for years. A Carnegie Mellon study of projects that adopted Cursor found a short-lived speed-up followed by lasting rises in code complexity and warnings.
- When your goal is to learn. In Anthropic's trial, people who delegated everything scored worst; those who asked the AI to explain kept their scores.
- Agents with access to production. Keep production credentials away from agents and enforce limits with permissions, not prompts.
A safer way to work
- Write a spec first. One user, three features, the rules that must never break. See the first app guide.
- Work in small steps. One feature per prompt, run it, then commit. A coding agent workflow makes this routine.
- Keep a rules file. An AGENTS.md stops the agent guessing.
- Read the diff and test the rules. Write tests for who can see what, and break a rule on purpose to prove a test fails.
- Separate development from production. Never give an agent the production database.
- Run the security checklist before you share the link.
- Use the AI to explain, not only to produce. Ask why, and fix some bugs yourself, so your own skills grow.
What it means for your career
- Employers are adapting, not uniform. Canva now tells engineering candidates to use AI in interviews and judges how they break down problems and whether they can find and fix mistakes in AI-generated code. Anthropic's candidate guidance says take-homes and live interviews are done without AI unless told otherwise. Read each company's rule.
- Entry-level pressure is real but not explained. A Stanford analysis of payroll data found a 13% relative employment decline for 22 to 25 year olds in the most AI-exposed jobs, and the authors say causality is not established.
- The skills that keep their value are the ones this page keeps returning to: specifying a problem, breaking it down, verifying results and owning quality.
- A portfolio that shows a live app, a spec, tests and a security check says more than one that shows only that you can prompt.
Frequently asked questions
What is vibe coding in simple words?
Describing an app to an AI in plain language and letting it write the code, then asking for changes until it works, often without reading the code. It is fast for prototypes and risky for anything that holds real users' data.
Who invented the term vibe coding?
Andrej Karpathy, in a post on X on 2 February 2025. Collins Dictionary later named it Word of the Year for 2025.
Is vibe coding the same as using GitHub Copilot or Cursor?
Not necessarily. Using an AI assistant while you read, test and understand the code is AI-assisted programming. Vibe coding, in the original sense, means accepting the output without reviewing it.
Is vibe coding safe?
Not on its own. Studies and incidents from 2025 and 2026 show missing access rules, exposed keys and weak security results when AI-written code is not reviewed. Reviewing the code, testing it and following a security checklist reduces the risk a lot.
Can a beginner learn to code by vibe coding?
It helps you start building, but research suggests learning drops if you delegate everything. Ask the AI to explain its code, read what it writes and fix some bugs yourself.
Will vibe coding replace programmers?
The evidence does not show that. Professionals mostly keep a human in the loop, and skills such as specifying, verifying and owning quality remain central. Entry-level hiring pressure is real, but studies say the cause is not established.
Learn this with mentors
Programmes for builders and engineers with admissions open: live classes, mentors and project work.
