Defense in Depth for Executing Model-Generated Code: Kernel-Isolated Sandboxes, Layered Network Policy, and Credential-Free Data Access on Multi-Tenant Kubernetes
International Journal of Research and Applied Innovations · 14.03.2025 · 8 (2)
DOI: 10.15662/ijrai.2025.0802017 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Research and Applied Innovations, online erschienen am 14.03.2025. Verfasst von Ramakrishna Kommineni.
Abstract (Originalsprache)
AI assistants that write and run code turn every prompt, and every document the model reads, into a possible source of executable instructions. This paper presents a defense-in-depth design for executing model-generated Python on a multi-tenant enterprise SaaS platform built on managed Kubernetes in Microsoft Azure, Amazon Web Services, and Google Cloud. It treats generated code as attacker-controlled and gives each attack path an independent control. The code sandbox runs under gVisor, a user-space kernel selected through a Kubernetes RuntimeClass. Sandbox traffic is governed by a four-layer network policy model that fails open in development, fails closed in production, and always allows DNS. Generated code never holds a credential: enterprise data is reached through Model Context Protocol connector services whose credentials flow from OpenBao through the External Secrets Operator, and Argo CD sync waves ensure that secrets exist before dependent AI services start. A runtime security agent adds independent detection. We map threats to controls, report lessons from operation, and define an evaluation methodology; quantitative results are not yet reported.
Zitieren
APA
Kommineni, R. (2025). Defense in Depth for Executing Model-Generated Code: Kernel-Isolated Sandboxes, Layered Network Policy, and Credential-Free Data Access on Multi-Tenant Kubernetes. International Journal of Research and Applied Innovations, 8(2). https://doi.org/10.15662/ijrai.2025.0802017
Vancouver
Kommineni R. Defense in Depth for Executing Model-Generated Code: Kernel-Isolated Sandboxes, Layered Network Policy, and Credential-Free Data Access on Multi-Tenant Kubernetes. International Journal of Research and Applied Innovations. 2025;8(2). doi:10.15662/ijrai.2025.0802017
BibTeX
@article{Kommineni2025,
author = {Kommineni, Ramakrishna},
title = {Defense in Depth for Executing Model-Generated Code: Kernel-Isolated Sandboxes, Layered Network Policy, and Credential-Free Data Access on Multi-Tenant Kubernetes},
journal = {International Journal of Research and Applied Innovations},
year = {2025},
volume = {8},
number = {2},
doi = {10.15662/ijrai.2025.0802017}
}
EUVIMED