Skip to content

NFTYoginis/vibe-code-audits

Repository files navigation

Vibe-Code Audit Teardowns

Empirical, named-stack audits of real apps shipped from Lovable, v0, Bolt.new, Cursor, and Replit Agent. Anonymized findings, severity-scored, free and MIT-licensed.


What this is

  • Stack-specific. Each teardown audits one named tool. Findings are filed under the stack that produced them, not against a generic AI-code checklist.
  • Empirical. Every claim cites a real audit. No theoretical advisories, no checklist-by-deduction.
  • Anonymized. App identities are redacted. Consent rules are documented and enforced before publication.

Browse teardowns

Every entry follows the same format: a verdict (shippable-as-is / shippable-after-fixes / scrap-and-restart), severity-ordered findings with anonymized code excerpts, a stack-pattern summary naming the tool defaults that produced the failures, and a net assessment of whether the issue is a tool problem, an operator problem, or both.

Stack pattern libraries

Each stack accumulates its own pattern library as audits land. A pattern claim requires at least three teardowns showing the same failure mode; until that threshold is met, observations stay in the individual teardowns. Pattern files name the tool default that produces the failure, the audits-in-evidence count, and the detection method an outside reader can use to spot it themselves.

Methodology

Audits run a fixed six-step process documented at methodology/how-we-audit.md. Severity follows the rubric at methodology/severity-rubric.md: CRITICAL means exploitable in production with a publicly-known technique, HIGH means exploitable with effort, MED means a production-quality miss, LOW means a code smell. Anonymization and consent rules are at methodology/how-we-anonymize.md. The methodology is reproducible — an outside auditor can run the same steps and submit findings.

Contribute

Outside auditors can submit full teardowns, single-finding additions to existing pattern files, and corrections to published findings. Consent and anonymization rules are non-negotiable. The PR template, severity-scoring guide, and review timeline are at CONTRIBUTING.md.

License

MIT — see LICENSE. Free to use, fork, and build on. Attribution appreciated; required for redistribution.

About

Empirical, named-stack, human-audited teardowns of code generated by Lovable, v0, Bolt.new, Cursor, Replit Agent. The specialist navigates the case files; it does not audit your code. Free under MIT.

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages