Sign InOpen Brain
arXivPaperNeeds Review

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

Kubernetes security patches improved sharply when the model received live service dependencies, showing that scanner findings alone omit context needed to avoid breaking workloads.

arXiv · Jul 28, 2026
Open Source Open MarkdownOpen JSON
Source Summary

KuTIE combines Istio call edges, Trivy findings, and service-account bindings before asking an LLM to patch Kubernetes configurations. Evaluation covered **248 trials** on a 36-deployment cluster with 31 findings across seven dependency classes.

Practical Implication

Feed remediation agents the runtime relationships their patches must preserve. For topology-dependent findings, correctness rose from **11.1% with scanner-only context** to **78.0% with topology context**.

Agent-Ready Context
KuTIE combines Istio call edges, Trivy findings, and service-account bindings before asking an LLM to patch Kubernetes configurations. Evaluation covered **248 trials** on a 36-deployment cluster with 31 findings across seven dependency classes.

Feed remediation agents the runtime relationships their patches must preserve. For topology-dependent findings, correctness rose from **11.1% with scanner-only context** to **78.0% with topology context**.

The topology-independent control showed **no improvement**, supporting a context-specific effect rather than generic prompt enrichment. The study used one purpose-built healthcare cluster, and one of seven dependency classes did not show the broader pattern.
Connected Context · Feed7 Judgment

KuTIE provides controlled evidence that context engineering should be dependency-aware, not merely additive: runtime topology sharply improved patches only when findings depended on that topology, while unrelated cases did not benefit. This strengthens the case for agents retrieving bounded, task-relevant relationships and for evaluating context effects per case, while limiting generalization to the tested cluster and dependency classes.

The Illusion of Robustness: Aggregate Accuracy Hides Prediction Flips under Task-Irrelevant ContextKuTIE supplies the positive counterpart to context-induced prediction flips: relevant topology helps dependent cases, reinforcing the need to measure context effects per task rather than through aggregate accuracy.AI tools for Forward Deployed Engineering — Vasuman Moza, Varick AgentsThe measured benefit of preserving real runtime relationships supports Varick’s broader prescription to capture how work actually operates before automating risky steps.RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AIKuTIE demonstrates why bounded evidence retrieval should select dependency relationships required by the patch instead of adding undifferentiated context.The AI bugpocalypse is here. Now what? - Jack Cable, CorridorIt gives a concrete implementation consequence for placing security review inside agent workflows: remediation agents need runtime dependencies, scanner findings, and identity bindings to avoid breaking relationships their patches must preserve.
Context Map
contextsecuritycoding#context-engineering#agent-reliability
Uncertainty
The topology-independent control showed **no improvement**, supporting a context-specific effect rather than generic prompt enrichment. The study used one purpose-built healthcare cluster, and one of seven dependency classes did not show the broader pattern.