Does Runtime Topology Context Improve LLM-Generated Kubernetes Security Patches? KuTIE Paper Review

Forum topic · 小凯 · 2026-07-30

Summary

A recent arXiv paper (2607.25995) introduces KuTIE (Kubernetes Topology Intelligent Engine), a system that improves LLM-generated Kubernetes security patches by conditioning generation on live cluster topology. Prior LLM-based remediation systems generate configuration patches from Kubernetes Security Posture Management (KSPM) findings in isolation, ignoring runtime service dependencies; fixes that are otherwise compliant can crash downstream callers or silently sever cross-cluster call edges. KuTIE builds runtime cluster context from Istio call edges, Trivy KSPM findings, and service account bindings read by workloads, then feeds this context to LLM patch generation. The system is evaluated on VulnCare, a purpose-built 36-deployment, four-namespace medical cluster with 31 injectable findings across seven dependency classes, labeled by topological dependency against cluster ground truth. Across 248 trials, topology context raised topology-dependent patch correctness from 11.1% to 78.0% (Δ=0.669), a gap holding across every model and six of seven dependency classes, ranging from credentials and network policies (Δ=0.95) to role-based access control (Δ=0.31); topology-independent controls showed no effect (Δ=0.0), isolating the result from generic prompt enrichment. The paper concludes that providing live service call graphs and their exposed service account bindings substantially improves remediation of topology-dependent findings compared to scanner-only context.

Paper Overview

Field: ML Authors: Farooq Shaikh Published: 2026-07-28 arXiv: 2607.25995

Abstract (original)

Kubernetes is central to the cloud-native ecosystem, orchestrating containerised workloads. Recent work suggests that large language models (LLMs) can automate cluster security remediation, generating configuration patches from Kubernetes Security Posture Management (KSPM) findings without human authoring. Such systems, however, prompt the model with each finding in isolation from the live service call graph, assuming general hardening knowledge suffices. This assumption breaks down whenever a patch must preserve a runtime service dependency invisible to the model: an otherwise compliant fix then carries a destructive functional blast radius, crashing downstream callers or silently severing call edges across the cluster.

Key Contributions

The paper introduces KuTIE (Kubernetes Topology Intelligent Engine), which builds live cluster context from three sources:

Istio call edges — runtime service-to-service traffic
Trivy KSPM findings — security posture issues
Service account bindings read by workloads

This runtime topology context is then used to condition LLM-based patch generation, so that fixes preserve real service dependencies the model would otherwise never see.

Evaluation: VulnCare Testbed

KuTIE is evaluated on VulnCare, a purpose-built medical cluster:

36 deployments, 4 namespaces
31 injectable findings across 7 dependency classes
Each finding labeled for topological dependency against cluster ground truth

Results

Across 248 trials:

| Metric | Value | |---|---| | Topology-dependent patch correctness (baseline) | 11.1% | | With topology context | 78.0% (Δ=0.669) | | Credentials / network policies | Δ=0.95 | | Role-based access control | Δ=0.31 | | Topology-independent controls | Δ=0.0 |

The improvement held for every model tested and for six of seven dependency classes. The null effect on topology-independent controls isolates the benefit from generic prompt enrichment.

Takeaway

Providing live service call graphs and their exposed service account bindings to LLMs significantly improves remediation of topology-dependent security findings compared to scanner-only context — reducing the destructive "functional blast radius" of otherwise compliant patches.

Paper: arXiv:2607.25995