EUVIMEDEUVIMED entdecken
Crossref

SAST IMPROVEMENTS USING LLM FOR CI/CD PIPELINES

V. Pavlyshyn, O. Krasko

Information and communication technologies, electronic engineering · 2026-06 · 6 (1) · 152-160

DOI: 10.23939/ictee2026.01.152 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Information and communication technologies, electronic engineering, 2026-04. Verfasst von V. Pavlyshyn und 1 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This paper proposes and evaluates a hybrid LLM-SAST pipeline for integrating Large Language Model (LLM)-based Static Application Security Testing (SAST) into Continuous Integration and Continuous Delivery (CI/CD) workflows of cloud-native telecommunication systems deployed on Kubernetes clusters. The rapid migration of 5G core network functions to containerized microservice architectures, where the Service-Based Architecture decomposes monolithic network elements such as Access and Mobility Management Function, Session Management Function, and Policy Control Function into independently deployable pods, dramatically expands the attack surface. This was demonstrated by the Salt Typhoon campaign that compromised at least nine major U.S. carriers in 2024, accessing metadata from over one million users and maintaining persistent access for up to three years before detection. Traditional rule-based SAST tools exhibit false positive rates exceeding 90 percent on standard benchmarks and miss critical business-logic vulnerabilities, creating an unsustainable triage burden in high-velocity telecom development environments. Recent research demonstrates that hybrid LLM-augmented SAST frameworks can improve vulnerability detection rates by a factor of 2.5 while reducing false positives by up to 91 percent. The proposed pipeline integrates deterministic rule-based static analysis engines such as Semgrep and CodeQL with LLM-powered semantic reasoning as a post-processing triage layer within a DevSecOps workflow. The rule-based engine produces structured findings including Common Weakness Enumeration identifiers and taint paths, which are then passed together with surrounding code context to the LLM for classification. The evaluation methodology employs the F1 score as the primary composite metric, computed as the harmonic mean of precision and recall, to ensure balanced assessment of detection completeness and noise reduction. Comparative analysis across nine tool configurations on standardized vulnerability benchmarks including the OWASP Benchmark with 2740 ground-truth test cases and CWE-Bench-Java with 120 validated vulnerabilities confirms that hybrid approaches achieve F1 scores of 0.91 to 0.95, significantly outperforming both traditional SAST tools with F1 of 0.10 to 0.55 and standalone LLM analysis with F1 of 0.61 to 0.68. Special attention is given to telecom-specific constraints including data sovereignty requirements under ETSI and 3GPP standards mandating self-hosted model deployment for air-gapped environments, policy-as-code enforcement through OPA Gatekeeper and Kyverno for compliance automation, software supply chain integrity through SBOM generation with cryptographic signing via Cosign and attestation verification through the SLSA framework, and Zero Trust principles implemented via service mesh for runtime protection of deployed network functions.

Zitieren

APA

Pavlyshyn, V., & Krasko, O. (2026). SAST IMPROVEMENTS USING LLM FOR CI/CD PIPELINES. Information and communication technologies, electronic engineering, 6(1), 152-160. https://doi.org/10.23939/ictee2026.01.152

Vancouver

Pavlyshyn V, Krasko O. SAST IMPROVEMENTS USING LLM FOR CI/CD PIPELINES. Information and communication technologies, electronic engineering. 2026;6(1):152-160. doi:10.23939/ictee2026.01.152

BibTeX
@article{Pavlyshyn2026,
  author = {Pavlyshyn, V. and Krasko, O.},
  title = {SAST IMPROVEMENTS USING LLM FOR CI/CD PIPELINES},
  journal = {Information and communication technologies, electronic engineering},
  year = {2026},
  volume = {6},
  number = {1},
  pages = {152--160},
  doi = {10.23939/ictee2026.01.152}
}
Kostenlos starten →