Why this exists

Vibe coding – building apps by describing what you want to AI tools like Cursor, Lovable, Bolt, and Replit – went from a niche technique to the default way non-technical founders ship a first product in about eighteen months. Y Combinator’s Garry Tan reported that roughly 25% of companies in the Winter 2025 batch had codebases that were 95%+ AI-generated. JetBrains’ 2025 developer survey of 24,534 developers across 194 countries found 85% now regularly use AI coding tools.

The tools are good enough that the resulting apps work in a demo. What they don’t guarantee is that the app is safe, scalable, or ready for real users – and a growing body of independent security research now quantifies exactly how large that gap is. This report collects what that research actually shows, not what vibe coding might hypothetically get wrong.

The headline numbers

45% of AI-generated code samples fail basic security tests.

Veracode tested over 100 large language models across 80 coding tasks in Java, Python, C#, and JavaScript. Results were worse in specific categories: 86% of samples failed to defend against cross-site scripting, and 88% were vulnerable to log injection. Java performed worst overall, at a 72% failure rate. A March 2026 re-test found the pass rate essentially unchanged since 2025, despite model vendors’ public claims of improved security-aware training.

CVE-2025-48757 – a critical Row-Level Security bypass in Lovable, CVSS 9.3.

Disclosed and logged in the National Vulnerability Database (patched by April 15, 2025), this vulnerability let unauthorized users bypass access controls entirely on affected Lovable-built apps. It’s one of 74+ CVEs now directly attributed to AI-generated code by Georgia Tech’s Vibe Security Radar project – a count that grew from 6 in January 2026 to 35 in March 2026 alone, with researchers estimating the true figure across the broader ecosystem at 5-10x higher.

2,038 critical vulnerabilities found across 1,400 scanned vibe-coded apps

Including more than 400 leaked secrets and 175 confirmed cases of exposed personal data – medical records, financial data, and authentication credentials, live and reachable in production. (Escape.tech, “State of Security of Vibe Coded Apps,” 2025 – apps built on Lovable, Base44, Bolt.new, and Vibe Studio.)

AI-assisted developers introduce security findings at roughly 10x the rate of their peers

Despite committing code only 3-4x faster. Privilege escalation paths in the resulting code rose 322%; architectural design flaws rose 153%. (Apiiro, scanning tens of thousands of Fortune 50 repositories, December 2024-June 2025.)

Nearly 80% of developers believe AI-generated code is more secure than code they write themselves

A belief the data above directly contradicts. (Snyk, 2025.)

Two failure modes worth naming specifically

1. Row-Level Security misconfiguration is the single most reproduced failure in AI-built apps.

This is the class of bug behind CVE-2025-48757: the AI tool correctly builds a working login flow and a working database, but never enforces who’s allowed to see whose data once both exist. It looks fine in a demo with one test account. It fails the moment a second real user signs up.

2. Secrets sprawl is accelerating specifically because of AI tooling, not despite it.

GitGuardian’s State of Secrets Sprawl 2026 report found 28.65 million new hardcoded secrets added to public GitHub in 2025 alone – a 34% year-over-year increase, the largest single-year jump on record. Secrets tied specifically to AI services grew 81% year over year. AI-assisted commits leak credentials at roughly double the rate of the platform-wide baseline. Popular MCP (Model Context Protocol) setup guides actively encourage the unsafe pattern – pasting API keys directly into config files – which is now spreading at ecosystem speed.

A third, newer risk: hallucinated dependencies

Independent USENIX Security 2025 research found that roughly 20% of AI-generated code samples reference software packages that don’t exist. Worse, 43% of those hallucinated names are reproduced consistently across similar prompts – predictable enough that attackers now pre-register the hallucinated package names before real developers install them, a technique researchers have named “slopsquatting.” One confirmed malicious package accumulated 30,000+ downloads in three months before detection. This is not yet a top-of-mind risk for most founders auditing their own vibe-coded app, and it should be.

What this means if you built your MVP with AI

None of this means AI coding tools were a bad choice – they clearly aren’t, or a quarter of a recent YC batch wouldn’t be running on them. It means the working demo and the production-ready app are two different engineering problems, and the gap between them is now well-documented, not speculative. The specific things worth checking before real users (or investors) arrive: Row-Level Security policies on every table that holds user data, hardcoded secrets anywhere in the repo or its config files, dependency names that don’t resolve to anything real, and whether your authentication flow was actually tested against a second, unauthorized account – not just your own.

That’s the exact scope of Beesoul’s free 3-day Vibe Code Audit.

Where this goes next (2027 outlook)

Three trends in the data above point the same direction. First, the vulnerability classes that are *growing* fastest – privilege escalation, architectural design flaws – are exactly the ones that require contextual judgment to catch, not pattern-matching; they will not be solved by the next model release the way syntax errors already have been (Apiiro’s data shows syntax errors in AI-generated code already dropped 76% even as architectural flaws rose 153% in the same period). Second, as vibe-coded MVPs increasingly carry real funding and real user data, technical due diligence on AI-generated codebases is moving from “nice to have” to a standard funding-round checkpoint – expect this to be explicit in term sheets within the next 12-18 months. Third, the attack surface is expanding past the code itself and into the tools: 2025 alone saw disclosed CVEs against Amazon Q, Cursor, and GitHub Copilot’s own rule-file processing, meaning the coding assistant is now part of what needs auditing, not just its output.

## Sources

– Veracode, [“GenAI Code Security Report”](https://www.veracode.com/blog/genai-code-security-report/), August 2025, and [“Spring 2026 GenAI Code Security Update”](https://www.veracode.com/blog/spring-2026-genai-code-security/), March 2026.

– NVD, [CVE-2025-48757](https://nvd.nist.gov/vuln/detail/CVE-2025-48757) (CVSS 9.3, Lovable Row-Level Security bypass), published May 29, 2025.

– Escape.tech, [“State of Security of Vibe Coded Apps”](https://escape.tech/state-of-security-of-vibe-coded-apps), 2025.

– Apiiro, [“4x Velocity, 10x Vulnerabilities”](https://apiiro.com/blog/4x-velocity-10x-vulnerabilities-ai-coding-assistants-are-shipping-more-risks/), September 2025.

– Cloud Security Alliance / Georgia Tech Vibe Security Radar, [“Vibe Coding’s Security Debt: The AI-Generated CVE Surge”](https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-generated-code-vulnerability-surge-2026/), April 2026.

– GitGuardian, [“The State of Secrets Sprawl 2026”](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/), March 2026.

– Spracklen et al., [“We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs”](https://www.usenix.org/system/files/conference/usenixsecurity25/sec25cycle1-prepub-742-spracklen.pdf), USENIX Security Symposium 2025.

– JetBrains, [“State of Developer Ecosystem 2025”](https://blog.jetbrains.com/research/2025/10/state-of-developer-ecosystem-2025/), October 2025.

– Cloud Wars / Snyk, [“Snyk’s AI Code Security Report Reveals Software Developers’ False Sense of Security”](https://cloudwars.com/cybersecurity/snyks-ai-code-security-report-reveals-software-developers-false-sense-of-security/), 2025.

– TechCrunch, [“A quarter of startups in YC’s current cohort have codebases that are almost entirely AI-generated”](https://techcrunch.com/2025/03/06/a-quarter-of-startups-in-ycs-current-cohort-have-codebases-that-are-almost-entirely-ai-generated/), March 2025.