Vibe coding: what an enterprise should allow, review and pay for

What vibe coding means, and what it does not

Karpathy described a way of working, not a product. He wrote that you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." He said he clicked "Accept All" and no longer read the diffs.¹

Merriam-Webster records the spelling "vibecoding" and the verb "vibe code." Its entry says the coder "does not need to understand how or why the code works."²

Developer Simon Willison drew the line in March 2025. He uses the term for "building software with an LLM without reviewing the code it writes." Code that a developer has reviewed, tested and can explain is, in his words, "not vibe coding, it's software development."³

Surveys use a looser meaning. Stack Overflow's 2025 Developer Survey defined vibe coding as generating software from LLM prompts. Even so, 72% of respondents said it was not part of their professional work, and another 5% said so emphatically.⁴

Where to allow vibe coding: prototypes versus production

The people who defined the practice also limited it. Karpathy scoped it to throwaway weekend projects. Willison's rule for production-quality work is that he won't commit code he couldn't explain to someone else.^1,3

The risk in production is not only bad code. In July 2025, Replit CEO Amjad Masad wrote that the company's coding agent, working in development, had "deleted data from the production database." He called it "unacceptable" and said Replit was rolling out automatic separation of development and production databases. Backups allowed a restore.⁵

Those cases suggest three zones:

• Prototypes, demos and personal tools: vibe coding is allowed, on public or synthetic data, with no other users depending on the result.¹
• Anything other people use or that holds company data: a named developer reviews and tests the code and can explain it before it ships.³
• AI agents with database or deployment access: no write path from a development session to production data.⁵

More on vibe coding

• AI Code Assistants & Coding AgentsBuyer Guide
• AI Code Review ToolsBuyer Guide
• Software Supply Chain Security & SBOMBuyer Guide
• Low-Code Application PlatformsBuyer Guide
• Codex vs Claude CodeComparison
• Cursor vs GitHub CopilotComparison
• Veracode vs SonarComparison
• RFP questions: ai model governanceRFP questions
• RFP questions: third party supply chainRFP questions
• State of Play: AI Governance & Responsible AIState of Play
• Vibe Coding: What It Is and What the Evidence Says for the EnterpriseArticle
• A One-Page AI Coding Policy for Regulated IndustriesArticle
• AI Coding Governance Policy BuilderApp
• AI Code Review GateFree download
• AI-Generated-Code Debt GuardrailsFree download
• Vibe coding (glossary)Glossary

Who may vibe code: developers and business staff

Platforms such as Lovable and Base44 generate whole apps from prompts and host them with their data. The app's security then depends on the generated code and on the platform.^6,7

Lovable's case shows the app layer failing. CVE-2025-48757 says an insufficient row-level security policy in Lovable through April 15, 2025 let remote unauthenticated attackers read or write arbitrary database tables of generated sites. MITRE scored it 9.3, Critical. Lovable disputes the CVE because each customer accepts responsibility for protecting its application's data.⁶

Base44's case shows the platform layer failing. Wiz Research reported in July 2025 that an attacker could create a verified account on private Base44 apps using only a non-secret app ID, bypassing controls including single sign-on. The flaw was fixed in less than 24 hours, and Wix, which had acquired Base44, said there was no evidence of past abuse.⁷

A vibe-coded app built by a business unit is therefore a third-party system holding company data. It needs the same vendor review as any SaaS purchase, and its access rules need testing by someone other than the platform.^6,7

Security risks of AI-generated code: insecure code, hallucinated packages, secrets

Veracode's 2025 GenAI Code Security Report, published July 30, 2025, gave more than 100 large language models 80 coding tasks. In 45% of test cases the models introduced a vulnerability from the OWASP Top 10. Java failed more than 70% of the time. The models failed to secure code against cross-site scripting in 86% of cases. Veracode scanned the output with its own static analysis product.⁸

Models also invent dependencies. A study presented at USENIX Security 2025 generated 576,000 code samples with 16 models. On average, at least 5.2% of the packages recommended by commercial models and 21.7% of those from open-source models did not exist. The authors counted 205,474 unique hallucinated package names.⁹

Attackers can register those names. OWASP's 2025 Top 10 for LLM applications describes attackers who find commonly hallucinated names and publish malicious packages under them. The OpenSSF calls this "slopsquatting."^10,11

Secrets leak as well. GitGuardian's 2025 State of Secrets Sprawl report found a 6.4% secret leakage rate in public repositories using Copilot, 40% higher than the average across all public repositories. That is a correlation in public code, not a measured effect of the tool.¹²

Some of those controls can be written into the assistant itself. The OpenSSF's guide to AI code assistant instructions, dated August 1, 2025, gives sample instructions such as "Do not add dependencies that may be malicious or hallucinated" and tells assistants to use the official package manager.¹¹

Does vibe coding make developers faster? What METR, DORA and Stack Overflow found

The controlled evidence and developers' impressions disagree. METR's randomized trial, published July 10, 2025, had 16 experienced developers work on 246 real issues from their own large open-source projects, with AI allowed on some issues and not on others.¹³

METR ran a follow-up from August 2025 with 57 developers, 143 repositories and more than 800 tasks. Its February 24, 2026 update estimates a speedup of -18% for the original developers, with a confidence interval of -38% to +9%. For new recruits the estimate is -4%, with an interval of -15% to +9%. METR calls the data "an unreliable signal." A growing share of developers said they would not want to do half their work without AI, and some held back tasks they did not want to do without it. METR believes developers are likely faster with AI tools in early 2026 than in early 2025, but says its data is only very weak evidence of how much.¹⁶

Google's DORA research program found in its 2025 report that AI adoption now has a positive relationship with software delivery throughput. It also found that AI adoption "does continue to have a negative relationship with software delivery stability." Its explanation is that more change volume, without strong automated testing and fast feedback, leads to instability. More than 80% of DORA's respondents said AI had raised their productivity, and 30% reported little or no trust in AI-generated code.^14,17

Stack Overflow's 2025 survey found more developers distrust the accuracy of AI tools (46%) than trust it (33%). Only 3% said they highly trust the output.¹⁸

Licensing, copyright and ownership of AI-generated code

In the US, code produced from prompts alone may not be protected by copyright. The US Copyright Office's January 2025 report on copyrightability concluded that "prompts alone do not provide sufficient human control to make users of an AI system the authors of the output." In vibe coding as Karpathy described it, the human accepts the output without reading it.^1,19

Licensing is the other exposure. GitHub says Copilot can generate code that matches code in public repositories. Depending on the organization's policy, matching suggestions are blocked or shown with the matching files and any detected licenses. GitHub says matches typically occur in less than one percent of inline suggestions.²⁰

What AI coding tools cost to buy and run

Business seats are priced per month, and heavy use is metered on top. As of November 2026, GitHub lists Copilot Business at $19 per granted seat per month with 1,900 AI credits per user, and Copilot Enterprise at $39 with 3,900. The credits go into a shared pool, and use beyond the pool costs $0.01 per credit.²¹

Anthropic lists Claude Enterprise at $20 per seat per month, billed annually, with a 20-seat minimum and usage charged at API rates. It says usage cost scales with model and task.²²

Review and debugging time come on top. In Stack Overflow's 2025 survey, 45% of respondents said debugging AI-generated code is more time-consuming. In METR's 2025 trial, the developers who believed AI had saved them time had taken 19% longer.^13,15