Security Analysis of LLM-Generated Web API Backends
TL;DR
A security assessment on 75 FastAPI backends generated by three contemporary LLMs revealed a disconnect between functional correctness and secure logic, which is interpreted as a review-risk pattern, which is called the human-in-the-loop paradox.
Abstract
The adoption of Large Language Models (LLMs) is changing how code is written, but the security implications of using LLMs to generate complete web API backends remain insufficiently characterized. Prior studies have assessed the security of LLM-generated code by detecting vulnerabilities in isolated code snippets; however, these assessments are too narrow to surface vulnerabilities that emerge in multi-endpoint backend services. To address this gap, we performed a security assessment on 75 FastAPI backends generated by three contemporary LLMs (Gemini 2.5 Pro, GPT 5.2, and DeepSeek V3.2). We designed a benchmark comprising the following five code generation categories: Authentication, Role-Based Access Control, File Uploads, Payment Processing, and Webhook Handling. Automated analysis of the generated APIs with Static Application Security Testing (SAST) tools, whose raw alerts we manually triaged to remove false positives, mainly revealed configuration problems, while Dynamic Application Security Testing (DAST) tools helped uncover runtime fragility such as application crashes on malformed input. Automated analysis was augmented with a manual assessment phase, which revealed vulnerabilities such as mass assignment and server-side request forgery. We also observed a disconnect between functional correctness and secure logic. In our trials, DeepSeek V3.2 produced the most vulnerable code despite the highest build success rate. Gemini 2.5 Pro had the highest startup-failure rate and produced shorter implementations, which coincided with fewer observed runtime findings among the APIs that started. GPT-5.2 fell between the two, characterized by near-correct security controls that took the right form but lacked the intended effect. We interpret this disconnect as a review-risk pattern, which we call the human-in-the-loop paradox: syntactically correct and well-structured LLM-generated code can still contain architectural vulnerabilities that are not revealed by surface-level inspection. The takeaway is that LLM-generated code warrants the same treatment as untrusted input from an external system.