Vibe Coding: The Promise and Risks of AI-Generated Code

Vibe coding is everywhere, but AI co-authored code shows 1.7x more major issues. What the quality data says and how to adopt it safely.

Vibe coding has spread fast, and it brings real risks to software quality and security. GitHub found in 2023 that 92% of US developers already used AI coding tools. Yet a CodeRabbit analysis found AI co-authored pull requests had 1.7 times more major issues than human-written code. Andrej Karpathy coined the term in February 2025, and a year later he was already promoting “agentic engineering” as the more disciplined successor. This guide covers the promise and the risks of vibe coding, what the quality data shows and how to adopt AI-assisted development responsibly.

92%
of US Developers Used AI Coding Tools (GitHub, 2023)
1.7x
More Major Issues in AI Co-Authored Code (CodeRabbit)
33%
of Developers Trust AI Accuracy (Stack Overflow 2025)

The Rise of Vibe Coding in Enterprise Development

Vibe coding means describing what you want in natural language and letting AI write the code. Karpathy described it as coding where you “fully give in to the vibes and forget that the code even exists.” Collins Dictionary named it Word of the Year for 2025. Since then, it has moved from weekend experiments into mainstream enterprise development.

Money has followed. The AI coding tools market is projected to reach $8.5 billion in 2026, and over $5 billion of venture capital went into AI coding tools in 2024 alone. Cursor also crossed $100 million in annual recurring revenue. So vibe coding is more than a passing trend, and the developer’s job is shifting from writing syntax to directing AI agents that handle implementation.

In Y Combinator’s Winter 2025 batch, a quarter of startups had codebases that were 95% AI-generated. Google has also said that more than a quarter of its new code is AI-generated. However, developer trust in AI accuracy fell from 43% in 2024 to 33% in 2025 in the Stack Overflow survey. As a result, the industry is using more of a technology it trusts less.

From Vibe Coding to Agentic Engineering

Karpathy proposed “agentic engineering” as the more mature successor to vibe coding in February 2026. The difference matters. Vibe coding means accepting AI output without full review. Agentic engineering, by contrast, is structured collaboration: AI agents handle implementation, while people provide architecture, review and quality assurance. In short, “forgetting the code exists” works for weekend projects, but it fails for production systems that run a business.

The Quality Problem Behind Vibe Coding

The data on vibe coding and AI-generated code quality is worrying, and every engineering leader should understand it before scaling AI-assisted development. Several independent studies point the same way: AI-generated code is faster to produce, but it brings more defects, more security issues and more maintenance work. Leaders who understand these trade-offs can keep the speed while adding the guardrails that protect quality.

1.7x More Major Issues
A CodeRabbit analysis of 470 open-source pull requests found that AI co-authored code had 1.7 times more major issues, including logic errors, wrong dependencies and flawed control flow. So AI-generated code needs more review effort than human-written code.
2.74x Cross-Site Scripting Flaws
Some security issues stand out. In the same CodeRabbit analysis, cross-site scripting flaws appeared 2.74 times more often in AI co-authored code. GitGuardian also counted 28.65 million new hardcoded secrets in public commits during 2025, and AI-assisted coding adds to that risk.
Technical Debt Is Rising
Forrester predicted that by 2026, 75% of technology decision-makers would face moderate to high technical debt, driven partly by fast AI-assisted development. So the speed gains from vibe coding can turn into maintenance costs that eat the early productivity benefits.
Comprehension Debt Crisis
Beyond technical debt, vibe coding creates “comprehension debt”: developers do not understand the code they are responsible for. Once the original prompts are lost, projects become very hard to maintain. As a result, the organization loses the ability to change its own systems.

The Security Risks of Vibe Coding at Scale

Vibe coding also creates security risks that traditional development practices were not built for. The scale of credential exposure alone calls for new ways to secure the software supply chain. The attack surface also runs both ways: AI coding tools have themselves become supply chain targets, with vulnerabilities disclosed in several major products.

Security Risk Scale Impact
Hardcoded Secrets 28.65M new secrets in public commits in 2025 ✗ Credential exposure keeps growing
Package Hallucinations About 20% of suggested packages did not exist in one study ✗ Attackers exploit through slopsquatting
Misconfigurations 75% more common in AI-generated code ◐ Traditional scanning catches most but not all
Tool Vulnerabilities CVEs disclosed against Cursor, Amazon Q, Copilot ◐ AI tools themselves are attack targets

In one study, about 20% of the packages suggested in AI-generated code samples did not exist. Attackers exploit this through “slopsquatting”, registering hallucinated package names as malware before developers install them. Existing SDLC frameworks, security training and CI/CD tools were also designed for human-written code, so they need deliberate updates for AI-specific failure patterns. Start by auditing which AI coding tools are in use, including ones developers added without IT approval.

The Productivity Paradox

METR ran a randomized controlled trial with experienced open-source developers working on their own codebases. The result was surprising: developers using AI tools took longer to finish tasks, yet they believed AI had sped them up. So the feeling of productivity and the measured result can point in opposite directions. In other words, quick wins on simple tasks can hide slower, lower-quality work on complex ones.

Talk to an ExpertAI-written code reaching production unchecked? Set up review and scanning with us.

How to Adopt AI-Assisted Development Without Catastrophic Risk

Smart organizations treat vibe coding tools as accelerators, not replacements, and give developers more time for architecture, system design and the hard problems AI still struggles with. AI works best for boilerplate, prototyping and routine tasks, where the speed is real and the risks are manageable. For security-critical paths and complex business logic, however, human engineering discipline is still essential. Systems that must be maintained for years need developers who understand every line they ship. So the “vibe and verify” workflow has become a recommended practice: generate code quickly with AI, then review, test and understand every line before it reaches production.

Safe Adoption Practices
Implementing mandatory AI code review before any merge to production
Running automated security scanning calibrated for AI failure patterns
Maintaining developer comprehension through “Vibe and Verify” workflows
Using AI for boilerplate and prototyping while hand-coding critical paths
Dangerous Practices
Accepting AI output without review for production codebases
Reducing junior engineer hiring based on AI productivity assumptions
Allowing unsanctioned AI tools without governance or audit trails
Measuring developer productivity by code volume rather than quality

Five Priorities for Governing Vibe Coding

Based on the quality and security data, here are five priorities for engineering leaders governing AI-assisted development:

  1. Audit all AI coding tools in use: Developers adopt tools without IT approval, so inventory every AI tool, including unsanctioned ones. That gives you the visibility governance and security need.
  2. Add AI-specific security scanning to CI/CD: AI co-authored code showed more vulnerabilities and 75% more misconfigurations in CodeRabbit’s analysis, so extend scanning to AI failure patterns. Also scan closely for hardcoded secrets in AI-assisted commits.
  3. Require human review for all AI-generated production code: Make sure reviews check that people understand the code, not just that it runs. As a result, comprehension debt does not build up quietly.
  4. Move from vibe coding to agentic engineering: Karpathy himself now promotes a more structured approach, so adopt one. AI handles implementation, while engineers own architecture and quality.
  5. Protect the junior engineer pipeline: Comprehension debt threatens resilience, so keep hiring and training junior engineers who build real understanding. They become the senior developers who can review AI output well later on. This matters for developer hiring in the AI era, where senior engineers stay scarce and fewer junior hires mean fewer seniors later.

Key Takeaway Vibe coding is everywhere: GitHub found 92% of US developers using AI coding tools back in 2023. However, AI co-authored code showed 1.7x more major issues, and cross-site scripting flaws appeared 2.74x more often. Forrester expected 75% of technology leaders to face moderate to high technical debt, GitGuardian counted 28.65M leaked secrets in 2025, and trust in AI accuracy fell to 33%. Karpathy now promotes agentic engineering. So audit AI tools, add AI-specific security scanning, require human review and protect junior developer pipelines.

Vibe coding: 92% of US developers used AI coding tools, 1.7x more major issues in AI co-authored code, 33% trust AI accuracy, and five priorities for engineering leaders
Vibe coding in three figures from GitHub, CodeRabbit and Stack Overflow, with five priorities.

Looking Ahead: From Vibe Coding to Agentic Engineering

Vibe coding is likely to mature into agentic engineering as the default professional workflow. AI agents will plan, write, test and ship code under structured human oversight. The developer’s job will move further from writing code to directing AI systems, while people own architecture, security and the quality decisions that keep systems reliable at scale.

Organizations that adopted AI-assisted development without governance, however, face years of technical and comprehension debt. Those that put structured review, security scanning and training in place from the start will move smoothly to agentic engineering. The advantage goes to organizations that capture AI’s speed without giving up code quality or human understanding.

For engineering leaders, AI code governance is about making sure productivity gains turn into maintainable, secure and understandable systems, instead of fragile codebases nobody can change when the AI tools fall short.

Our ServiceDevOps and IT Automation

Frequently Asked Questions

Frequently Asked Questions
What is vibe coding?
Vibe coding is a way of developing software where you describe what you want in natural language and AI writes the code. Andrej Karpathy coined the term in February 2025, and Collins named it Word of the Year. GitHub found 92% of US developers using AI coding tools in 2023, and the market is projected to reach $8.5B in 2026.
How much riskier is AI-generated code?
In CodeRabbit’s analysis, AI co-authored code had 1.7x more major issues and 75% more misconfigurations, and cross-site scripting flaws appeared 2.74x more often. In one study, about 20% of suggested packages did not exist, which attackers exploit through slopsquatting. AI coding tools themselves have also had security vulnerabilities.
What is comprehension debt?
Comprehension debt means the people responsible for code do not understand what it does. Traditional technical debt means messy code. Comprehension debt is worse, because developers cannot debug, maintain or change systems they did not write or review. If AI tools become unavailable, nobody can change those systems.
What is agentic engineering?
Agentic engineering is the more structured successor to vibe coding that Karpathy promoted in February 2026. AI agents handle implementation, while people provide architecture, review and quality assurance. The key difference is more oversight and review instead of blindly accepting AI output.
Should enterprises ban vibe coding?
No. Banning AI coding tools is impractical when most developers already use them. Instead, put governance in place: mandatory review, AI-specific security scanning, approved tool lists and structured workflows. Use AI for boilerplate and prototyping, and hand-code security-critical paths.

References

  1. 1.7x Issues, 2.74x XSS, 75% Misconfigurations, Trust Decline: Prof. Hung-Yi Chen — The Dark Side of Vibe Coding: The AI Code Quality Crisis
  2. 28.65M Secrets, Slopsquatting, AI Tool Vulnerabilities: CSA — Vibe Coding Security Crisis: Credential Sprawl and SDLC Debt
  3. METR Trial, $8.5B Market, Agentic Engineering, Karpathy Evolution: Hashnode — The State of Vibe Coding in 2026: Adoption Won, Now What?
  4. 92% of US Developers Use AI Coding Tools (2023): GitHub Blog — Survey Reveals AI’s Impact on the Developer Experience
  5. Trust in AI Accuracy Fell to 33%: Stack Overflow — 2025 Developer Survey: AI
  6. Randomized Trial With Experienced Developers: METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Newsletter

Security Insights, Delivered.

Practical, vendor-neutral analysis of threats, tools, and architecture decisions.

Book a Consultation

A practice principal will contact you within 1 business day.

Thank you! A Signisys principal will be in touch within 1 business day.