A Comparative Study of AI-Generated and Human-Written Kubernetes Manifests

Kubernetes is susceptible to security and quality issues when misconfigured; meanwhile, the use of generative models such as ChatGPT to produce manifests is rising. This study compares the configuration quality of human-authored manifests (۱۹۸۵ open-source files) with a published ۲۰۲۳ corpus of AI-generated manifests (۹۸ ChatGPT-generated files aligned with a prior MSR study). Using an aligned KubeLinter-based framework, we analyze six configuration smells and report density (occurrences per ۱۰۰۰ raw lines) and proportion (percentage of files affected). We also add a balanced-size robustness check that subsamples ۹۸ human-written manifests ۱۰۰۰ times. The results show that missing CPU and memory specifications remain more prevalent in the AI corpus under this balanced comparison. In contrast, omissions in secure execution, such as missing runAsNonRoot and readOnlyRootFilesystem, are important in both corpora. An issue-level smell-by-kind analysis of the human corpus shows that ۸۵.۱۶% of target-smell occurrences concentrate in Deployment and StatefulSet resources. The findings provide a focused empirical baseline and support the use of human review, static analysis, and policy enforcement before integration or deployment.

کلیدواژه ها:

Cloud-Native Security ، Configuration Smells ، Container Orchestration ، Large Language Models ، Static Analysis