Vibe coding is building software by telling an AI model what you want and accepting the code it produces without reviewing it. Merriam-Webster defines it as "writing computer code in a somewhat careless fashion, with AI assistance." Andrej Karpathy coined the term in a post on X on February 2, 2025.
Where the term "vibe coding" came from
Karpathy described a way of working, not a product. "There's 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," he wrote in the original post. In the same post he wrote: "I 'Accept All' always, I don't read the diffs anymore." He also set its scope: "It's not too bad for throwaway weekend projects, but still quite amusing."
Merriam-Webster's entry records the spelling "vibecoding" and the verb "vibe code." It credits Karpathy, whom it describes as "a co-founder of OpenAI." It says that in vibe coding "the coder does not need to understand how or why the code works, and often will have to accept that a certain number of bugs and glitches will be present."
Collins named vibe coding its Word of the Year 2025 on November 6, 2025. Its definition is broader: "the use of artificial intelligence prompted by natural language to write computer code."
Is all AI-assisted coding vibe coding?
No. The definitions split on review. Karpathy's post and Merriam-Webster describe code that nobody reads. Collins covers any code written from natural-language prompts.
Developer Simon Willison drew the line in March 2025: "When I talk about vibe coding I mean building software with an LLM without reviewing the code it writes." His rule for production code is that he won't "commit any code to my repository if I couldn't explain exactly what it does to somebody else." In the same post he wrote that reviewed, tested code a developer can explain is "not vibe coding, it's software development."
Surveys use the broader meaning. The Stack Overflow 2025 Developer Survey defined vibe coding as "the process of generating software from LLM prompts." Of its 26,564 respondents to that question, 72.2% answered "No" when asked whether vibe coding was part of their professional development work, per the survey's AI section. Another 5.3% answered "No, emphatically," and 11.9% answered "Yes," in the same results.
Most of the studies below measure AI output or AI-assisted work in general. Vibe coding is the case where that output ships without a human review in between.
What the security research shows
Veracode published its 2025 GenAI Code Security Report on July 30, 2025. It gave more than 100 large language models 80 code-completion tasks with known potential for weaknesses, then scanned the output with Veracode Static Analysis.
- In 45 percent of all test cases, the models introduced vulnerabilities in the OWASP Top 10, per Veracode.
- Given a choice between a secure and an insecure way to write the code, the models "chose the insecure option 45 percent of the time," the release states.
- Java had a security failure rate over 70 percent, and Python, C# and JavaScript had rates between 38 and 45 percent, per Veracode.
- The models failed to secure code against cross-site scripting (CWE-80) in 86 percent of cases and against log injection (CWE-117) in 88 percent, per the same release.
- Syntax improved across model releases while security performance stayed "unchanged over time," and larger models did not "perform significantly better than smaller models," according to Veracode.
The tasks were narrow completions, not whole applications, and the scanner is Veracode's own product. The release recommends Veracode tools for remediation. Veracode CTO Jens Wessling tied the result to vibe coding: developers "do not need to specify security constraints to get the code they want, effectively leaving secure coding decisions to LLMs," he said in the release.
What the quality research shows
Developers report that AI output needs correction. In the Stack Overflow 2025 survey, 66% of respondents to the question on frustrations chose "AI solutions that are almost right, but not quite," from 31,476 responses. In the same question, 45.2% chose "Debugging AI-generated code is more time-consuming." More respondents distrust the accuracy of AI tools (46%) than trust it (33%), and 3% "highly trust" the output, per the survey.
Google's DORA research program surveyed 4,867 respondents for its 2025 State of AI-assisted Software Development report. Its executive summary (page 4) states: "AI adoption now improves software delivery throughput, a key shift from last year. However, it still increases delivery instability." DORA's report page describes AI's primary role as "an amplifier, magnifying an organization's existing strengths and weaknesses."
GitClear examined 211 million changed lines from 2020 to 2024 in repositories owned by Google, Microsoft, Meta and other companies. It reports a spike in duplicate code blocks, more short-term churn and a continued decline in moved lines, which it treats as a measure of code reuse. Its summary states that "copy/paste" exceeded "moved" code "for first time in history."
What the productivity research shows
METR ran a randomized controlled trial published July 10, 2025. It recruited 16 experienced developers from large open-source repositories averaging more than 22,000 stars and 1 million lines of code. Each of 246 real issues was randomly assigned to allow or disallow AI. The tools used were primarily Cursor Pro with Claude 3.5 and 3.7 Sonnet.
- With AI allowed, developers took 19% longer to complete issues, per METR.
- Before the study, developers expected AI to speed them up by 24%, according to METR.
- After it, they still believed AI had sped them up by 20%, the study page reports.
METR states that it does not claim its developers or repositories "represent a majority or plurality of software development work."
METR published follow-up data on February 24, 2026, from 57 developers across 143 repositories and more than 800 tasks, starting in August 2025. For developers from the original study it estimates "a speedup of -18% with a confidence interval between -38% and +9%," and for new recruits "-4%, with a confidence interval between -15% and +9%," per the update. METR says this data gives "an unreliable signal." Developers increasingly declined to take part because they did not want to work without AI, which METR says "likely biases downwards our estimate of AI-assisted speedup." It believes developers are likely "more sped up from AI tools now" than in early 2025, and calls its data "only very weak evidence for the size of this increase."
Self-reports run higher. Over 80% of DORA's 2025 respondents said AI enhanced their productivity, according to Google's summary of the report, which also puts AI adoption among software professionals at 90%. METR's 2025 trial is the controlled case where developers' estimates and measured time pointed in opposite directions.
Which incidents show the risks?
A Replit agent deleted production data in July 2025
Jason Lemkin was building an app on Replit and posting about it as "Vibe Coding Day 7." On July 18, 2025 he posted: "@Replit goes rogue during a code freeze and shutdown and deletes our entire database."
Replit CEO Amjad Masad responded on July 20, 2025: "@Replit agent in development deleted data from the production database. Unacceptable and should never be possible." He said Replit had started "rolling out automatic DB dev/prod separation" and was building staging environments. He said backups allowed "a one-click restore for your entire project state." He said Replit was working on "a planning/chat-only mode" after the "code freeze" complaint, and would refund Lemkin and run a postmortem.
The agent working in development could write to production. An instruction to freeze code did not stop it.
Lovable-generated apps exposed their databases (CVE-2025-48757)
The NVD entry for CVE-2025-48757 reads: "An insufficient database Row-Level Security policy in Lovable through 2025-04-15 allows remote unauthenticated attackers to read or write to arbitrary database tables of generated sites." The same entry notes the CVE "is disputed by the Supplier because each individual customer of the Lovable platform accepts a responsibility over protecting the data of their application." It carries a CVSS 3.1 base score of 9.3, Critical, from MITRE, per NVD.
The researcher, Matt Palmer, worked at Replit, a competing platform, per his disclosure. His statement says a scan completed March 21, 2025 "identified 303 endpoints across 170 projects (approximately 10.3% of the 1645 analyzed) with inadequate RLS settings." It says Lovable's later security scanner "merely checks for the existence of any RLS policy, not its correctness."
A platform flaw opened private apps on Base44 in July 2025
Wiz Research reported on July 29, 2025 that an attacker could create a verified account for private Base44 applications "by providing only a non-secret app_id value to undocumented registration and email verification endpoints." Wiz says this bypassed the platform's authentication controls, "including Single Sign-On (SSO)." Wiz describes Base44 as "recently acquired by Wix." Per Wiz, the flaw was fixed in less than 24 hours, and Wix confirmed there was no evidence of past abuse.
The three cases fail at different layers: the agent's permissions, the generated app's access rules, and the platform hosting the apps.
Where vibe coding belongs in an enterprise
The research and the incidents point to one boundary. Vibe coding fits work where a failure costs little. It does not fit anything that holds real data or reaches production. Karpathy's own scope was throwaway weekend projects.
- Use it for prototypes, demos and personal tools on public or synthetic data. No one else depends on the result, and no customer data is at stake.
- Require review before anything ships to other users. Code that a developer has reviewed, tested and can explain is no longer vibe coding under Willison's definition. Veracode's results describe model output before that review.
- Give development agents no write path to production data. Replit's fix was dev/prod separation. The code freeze Lemkin described was an instruction, and the agent deleted the data anyway.
- Treat vibe-coding platforms as third-party vendors. They host the apps and the data. The Base44 flaw was in the platform, and Lovable's dispute of its CVE puts data protection on the customer.
- Do not accept a platform's built-in scanner as a security review. Check access rules such as row-level security with your own tests.
- Measure delivery outcomes, not developer estimates. METR found a gap between how fast developers felt and how fast they were. DORA measures instability with change fail rate and deployment rework rate, which count failed and unplanned deployments.
For the policy itself, with zones, data rules, intake and ownership, see A One-Page AI Coding Policy for Regulated Industries.
Independent. No sponsorships. Unsubscribe anytime.