Misconfigured by Default: The Prevalence, Validity, and Tool Coverage of Kubernetes Misconfigurations in the Wild

Kubernetes manifests are the de facto interface to production infrastructure, with permissive defaults: a workload specifying nothing runs with an automounted token, no resource limits, and no explicitly configured seccomp profile. We measure configuration in 151 Helm charts and 60 public GitHub repositories (16,243 documents, 2,811 workloads) with a 25-rule engine. We validated the engine on a 457-finding stratified sample: AI-generated pre-labels were author-checked — the 92 protocol-sensitive rows adjudicated, a blinded 30-row subset (author–AI agreement: kappa = 0.526 [0.17, 0.84]) — the rest retain accepted pre-labels, 16 corrected by a blinded re-labeling audit. Population-weighted precision is 0.676 [0.60, 0.81] (cluster bootstrap by source): ~18,600 [16,500, 22,100] of 27,460 raw findings are estimated genuine. Precision is rule-dependent: absence checks (seccomp, privilege escalation, root prevention) showed no false positives in the validation sample, while context-sensitive rules (committed secrets, host namespaces, probes) are mostly benign-by-design. The defaults dominate: tokens automount in 88.6% of production GitHub workloads, CPU limits — a contested hardening practice — are missing in 75.3%; 1.4% run a privileged container. Helm charts beat raw manifests on most practices, but by only 1.3–1.7x in workload-level prevalence. We also audit the auditors: on identical input, four scanners cover at most 16 of the 25 practices, diverging up to 1.4x on the same practice through counting semantics. Scored against reference labels (Helm arm, n = 126), they effectively detect 51–72% of the engine's reference-positive findings and flag 84–95% of reference-benign findings where a check exists, under our contextual protocol. Manuscript submitted to Empirical Software Engineering (2026). Replication package (data and code): https://doi.org/10.5281/zenodo.23969482

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24297703
Primary Topic
Software System Performance and Reliability
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Misconfigured by Default: The Prevalence, Validity, and Tool Coverage of Kubernetes Misconfigurations in the Wild

Girish Mallya
3 links
Zenodo (CERN European Organization for Nuclear Research)
Software System Performance and Reliability
preprint

Misconfigured by Default: The Prevalence, Validity, and Tool Coverage of Kubernetes Misconfigurations in the Wild

Girish Mallya
preprint en

Abstract

Kubernetes manifests are the de facto interface to production infrastructure, with permissive defaults: a workload specifying nothing runs with an automounted token, no resource limits, and no explicitly configured seccomp profile. We measure configuration in 151 Helm charts and 60 public GitHub repositories (16,243 documents, 2,811 workloads) with a 25-rule engine. We validated the engine on a 457-finding stratified sample: AI-generated pre-labels were author-checked — the 92 protocol-sensitive rows adjudicated, a blinded 30-row subset (author–AI agreement: kappa = 0.526 [0.17, 0.84]) — the rest retain accepted pre-labels, 16 corrected by a blinded re-labeling audit. Population-weighted precision is 0.676 [0.60, 0.81] (cluster bootstrap by source): ~18,600 [16,500, 22,100] of 27,460 raw findings are estimated genuine. Precision is rule-dependent: absence checks (seccomp, privilege escalation, root prevention) showed no false positives in the validation sample, while context-sensitive rules (committed secrets, host namespaces, probes) are mostly benign-by-design. The defaults dominate: tokens automount in 88.6% of production GitHub workloads, CPU limits — a contested hardening practice — are missing in 75.3%; 1.4% run a privileged container. Helm charts beat raw manifests on most practices, but by only 1.3–1.7x in workload-level prevalence. We also audit the auditors: on identical input, four scanners cover at most 16 of the 25 practices, diverging up to 1.4x on the same practice through counting semantics. Scored against reference labels (Helm arm, n = 126), they effectively detect 51–72% of the engine's reference-positive findings and flag 84–95% of reference-benign findings where a check exists, under our contextual protocol. Manuscript submitted to Empirical Software Engineering (2026). Replication package (data and code): https://doi.org/10.5281/zenodo.23969482

Zenodo (CERN European Organization for Nuclear Research)
Software System Performance and Reliability
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon