Assessing the Security of “Vibe Coding”: An Analysis of Vulnerabilities in Web Applications Generated by LLMs
Abstract
Vibe coding shifts validation from source code review to observable behaviour, which may allow vulnerabilities that do not affect perceived functionality to go unnoticed. This paper compares the vulnerability surface of four e-commerce web applications generated by four distinct LLMs from a single specification without additional security requirements. Each application was scanned twice with OWASP ZAP 2.17.0, once per access profile, under identical configuration, and complemented by a manual audit of access control and session management. The scans reported between 32 and 58 unique findings per model and up to nine high-severity findings. Manual reproduction refuted all 18 high-severity findings, the most pronounced difference between models. The defects common to all four applications and not reported by the scanner concern the session lifecycle: tokens remain valid after logout, and signing secrets are embedded in the source code as default values.