What is vibe coding? Meaning, origin and security risks

Describing the outcome, not the instructions

Programming, traditionally, means writing precise instructions a machine carries out step by step: lines of code in a language like the one covered in the article on PHP, thought through and checked by the person writing them. Vibe coding turns that around: instead of writing the code, you describe it. You tell an artificial intelligence model what you want — “add a button that exports the data to PDF,” “shrink the padding on the left of the sidebar” — and let the model translate that request into working code, often running it right away.

The defining trait, by the term’s original meaning, isn’t so much using an AI to write code — that became common practice between 2025 and 2026 — but giving up on checking it. Someone doing vibe coding accepts the changes the model proposes without rereading the “diffs,” the line-by-line comparison between the old and new version of the code. What gets judged is the result on screen, not how it got there.

Vibe coding versus classic AI-assisted coding

Not every use of artificial intelligence in programming is vibe coding. Tools like GitHub Copilot, which suggest lines or functions as you type, or general-purpose chatbots used to get unstuck on a specific problem, have been around for years: MIT Technology Review calls this classic “AI-assisted coding,” and what separates it from vibe coding mostly comes down to the control that stays with the person programming.

Practical example: a developer who uses Copilot to complete a function they already designed, then checks its logic line by line, is doing classic AI-assisted coding. A developer who says out loud “add a cart with an automatic discount above 50 euros” and accepts whatever code appears without checking it, trusting that the app seems to work, is doing vibe coding by the term’s original definition.

From a tweet to word of the year

The term comes from a post published on X (Twitter) by Andrej Karpathy, an OpenAI co-founder and former director of artificial intelligence at Tesla, on 2 February 2025. Karpathy described programming by relying almost entirely on an AI assistant built into a code editor, even dictating requests out loud with a transcription app. As he put it himself, “I don’t read the diffs anymore.” The post passed 4.5 million views according to converging secondary sources, and a year later Karpathy himself called it a throwaway thought he’d fired off, saying he could never predict which of his posts would take off.

From there the term spread fast: within a few months, articles, guides and discussions on the topic were everywhere, reaching a level of recognition that usually takes years, not months. On 6 November 2025 Collins Dictionary named it Word of the Year 2025, defining it as “the use of artificial intelligence prompted by natural language to write computer code.” The same shortlist of AI-themed words Collins considered for 2025 also included “clanker” and “broligarchy”; around the same time Merriam-Webster instead picked “slop” as its own word of the year, signaling how central the topic of AI-generated content had become in that period’s public debate.

Who benefits, and who’s most at risk

According to Tobin South, a researcher at the MIT Media Lab, vibe coding mostly benefits two groups at opposite ends of the spectrum: experienced programmers, who can quickly spot and fix a mistake the AI generates, and absolute beginners who have a clear idea of what they want to build but wouldn’t know how to translate it into code on their own. The sources don’t say explicitly who benefits least, but it’s reasonable to guess it’s people still learning to program: understanding code, which is also how you learn to write better code, gets handed over almost entirely to the model in that case. The reasoning behind these systems is the same one that lets a model recognize patterns across huge amounts of data, as covered in the article on machine learning, and that more broadly powers generative AI, capable of producing text, images or code from a plain-language instruction.

The documented risks: when “working” doesn’t mean “secure”

One of the risks discussed most often by industry studies is precisely the lack of human review. One figure stands out from the rest: in Carnegie Mellon University’s SusVibes benchmark, run on 200 tasks built from 108 open-source Python projects, 61% of the code generated by an AI agent built on Claude Sonnet 4 passed functional tests, but only 10.5% also passed a security review. In practice, out of ten features that appear to work, roughly nine carry at least one exploitable vulnerability.

It isn’t an isolated case. Veracode tests across more than 100 language models found that 45% of development tasks handed to an AI introduce at least one security vulnerability, often falling into the OWASP Top 10 categories, the reference standard for the most common flaws in web applications. The Cloud Security Alliance also flags another, less visible but widespread risk: credential leaks. In 2025, 28.65 million secrets — API keys, passwords, access tokens — were found mistakenly left in code published on GitHub, a 34% increase from the year before, the sharpest single-year jump on record. Faster development and publishing cycles enabled by AI are likely among the factors behind that jump, though pinning the entire increase on vibe coding alone would be an oversimplification: the phenomenon is tied more broadly to the growing use of AI tools in software development.

New classes of attack have also appeared with no real equivalent in code written entirely by human programmers. “Slopsquatting,” for instance, exploits the fact that AI models sometimes recommend nonexistent software package names, “hallucinated” during generation: across a sample of 576,000 cases analyzed by the Cloud Security Alliance, 20% contained invented packages. Attackers register those names on public repositories like npm or PyPI, waiting for someone to install them by mistake and unknowingly bring malicious code into their project.

The gap between trust and quality

Data collected by the Cloud Security Alliance also describes a gap between how much developers trust these tools and how well they actually perform in practice. 91% of surveyed developers use AI coding tools, and about 22% of code merged into projects is written by artificial intelligence. Over the same period, though, developer trust in AI accuracy fell from 40% to 29% year over year, and 66% say they spend more time fixing AI-generated code that looks almost right but isn’t — a sign that more use of these tools hasn’t come with more trust in how reliable they actually are. On a related front, those working on regulating the use of artificial intelligence — as covered in the article on the EU AI Act, Europe’s AI regulation — are paying growing attention to exactly the systems that can affect people’s safety and data, a category that also includes applications built entirely through vibe coding and put into production without extra checks.