Selected work

Representative Papers

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

Jul 28, 2026

This work addresses the challenge that large language models (LLMs) often generate Kubernetes security patches without awareness of runtime topological context, risking service dependency violations and functional disruptions. To overcome this limitation, the authors propose KuTIE, a novel framework that systematically integrates runtime topology information—including Istio call graphs, Trivy security findings, and service account bindings—into LLM prompt engineering to enable context-aware automated remediation. Experimental evaluation on the VulnCare platform demonstrates that incorporating such topological context substantially improves the correctness of dependency-sensitive patches, increasing accuracy from 11.1% to 78.0%. These results validate the effectiveness and innovation of the proposed approach in enhancing the precision of LLM-generated security configurations for Kubernetes environments.

0 citationsRead paper

A Comparison of Kubernetes Compliance Standards and Configuration Scanners

Jun 23, 2026

This study addresses the significant inconsistencies among current Kubernetes security hardening guidelines and configuration scanning tools in terms of recommendation coverage and risk scoring, which hinder effective security configuration decisions. The authors systematically analyze eight widely adopted hardening guides to derive a unified benchmark of 79 configuration recommendations and conduct a structured empirical evaluation of ten static scanning tools. For the first time, they establish a standardized framework for assessing Kubernetes configuration security. Their findings reveal substantial discrepancies in both coverage and risk assessment across existing guidelines and tools, underscoring the urgent need for a transparent and consistent security evaluation methodology. This work provides the community with a reproducible benchmark and actionable criteria for tool selection and policy alignment.

0 citationsRead paper

BlobShuffle: Cost-Effective Repartitioning in Stream Processing Systems via Object Storage Exemplified with Kafka Streams

Jun 02, 2026

This work addresses the high network costs and operational overhead incurred by cross-availability-zone repartitioning in large-scale stream processing systems. It introduces, for the first time, cloud object storage into the stream shuffle pipeline through a pluggable architecture built atop Kafka Streams: upstream operators batch-write repartitioning data to object storage and emit lightweight notifications, while downstream operators fetch data on demand. By integrating configurable batching and distributed caching, the approach ensures consistency without requiring modifications to Kafka or underlying infrastructure. Experimental evaluation on AWS Kubernetes clusters demonstrates over 40× reduction in repartitioning costs, 95th-percentile latency under 2 seconds, and throughput exceeding 2 GiB/s, substantially enhancing scalability and efficiency.

0 citationsRead paper

FORGE: Multi-Agent Graduated Exploitation and Detection Engineering

Jun 02, 2026

This work addresses the growing gap between the surge in vulnerability disclosures and organizations’ capacity to assess them, exacerbated by the longstanding fragmentation in research on exploit generation, prioritization, and detection rule engineering. To bridge this divide, the authors propose FORGE, a multi-agent system that introduces “exploit depth tiers” as a unifying framework for end-to-end vulnerability assessment. FORGE constructs targeted applications from CVE metadata, employs LLM-guided four-tier evaluation to iteratively refine exploits, and automatically generates Sigma and Snort detection rules from OpenTelemetry behavioral traces. Evaluated on 603 CVEs, FORGE achieves a 67.8% end-to-end success rate for L1+ exploits at a cost of $1.50 per CVE, produces L2+ detection rules that significantly improve normalized coverage (p=0.035), and yields Snort rules with zero false positives on benign traffic in 93.4% of cases.

0 citationsRead paper

Navigating the Sea of LLM Evaluation: Investigating Bias in Toxicity Benchmarks

May 11, 2026

Current benchmarks for evaluating toxicity in large language models exhibit underappreciated systematic biases that may lead to the deployment of unsafe models. This work systematically investigates how variations in task formulation—such as text completion versus summarization—input data domains, and evaluated models interact with multiple toxicity metrics. It reveals, for the first time, that both task type and data domain significantly influence toxicity scores. Experiments demonstrate that existing benchmarks are prone to misclassifying content as harmful when tasks are altered and show inconsistent performance across domains, highlighting their fragility and dependence on specific model-task configurations. These findings underscore the urgent need for more robust and reliable toxicity evaluation frameworks.

0 citationsRead paper

Recent publications

Latest Papers

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

Jul 28, 2026

This work addresses the challenge that large language models (LLMs) often generate Kubernetes security patches without awareness of runtime topological context, risking service dependency violations and functional disruptions. To overcome this limitation, the authors propose KuTIE, a novel framework that systematically integrates runtime topology information—including Istio call graphs, Trivy security findings, and service account bindings—into LLM prompt engineering to enable context-aware automated remediation. Experimental evaluation on the VulnCare platform demonstrates that incorporating such topological context substantially improves the correctness of dependency-sensitive patches, increasing accuracy from 11.1% to 78.0%. These results validate the effectiveness and innovation of the proposed approach in enhancing the precision of LLM-generated security configurations for Kubernetes environments.

0 citationsRead paper

A Comparison of Kubernetes Compliance Standards and Configuration Scanners

Jun 23, 2026

This study addresses the significant inconsistencies among current Kubernetes security hardening guidelines and configuration scanning tools in terms of recommendation coverage and risk scoring, which hinder effective security configuration decisions. The authors systematically analyze eight widely adopted hardening guides to derive a unified benchmark of 79 configuration recommendations and conduct a structured empirical evaluation of ten static scanning tools. For the first time, they establish a standardized framework for assessing Kubernetes configuration security. Their findings reveal substantial discrepancies in both coverage and risk assessment across existing guidelines and tools, underscoring the urgent need for a transparent and consistent security evaluation methodology. This work provides the community with a reproducible benchmark and actionable criteria for tool selection and policy alignment.

0 citationsRead paper

BlobShuffle: Cost-Effective Repartitioning in Stream Processing Systems via Object Storage Exemplified with Kafka Streams

Jun 02, 2026

This work addresses the high network costs and operational overhead incurred by cross-availability-zone repartitioning in large-scale stream processing systems. It introduces, for the first time, cloud object storage into the stream shuffle pipeline through a pluggable architecture built atop Kafka Streams: upstream operators batch-write repartitioning data to object storage and emit lightweight notifications, while downstream operators fetch data on demand. By integrating configurable batching and distributed caching, the approach ensures consistency without requiring modifications to Kafka or underlying infrastructure. Experimental evaluation on AWS Kubernetes clusters demonstrates over 40× reduction in repartitioning costs, 95th-percentile latency under 2 seconds, and throughput exceeding 2 GiB/s, substantially enhancing scalability and efficiency.

0 citationsRead paper

FORGE: Multi-Agent Graduated Exploitation and Detection Engineering

Jun 02, 2026

This work addresses the growing gap between the surge in vulnerability disclosures and organizations’ capacity to assess them, exacerbated by the longstanding fragmentation in research on exploit generation, prioritization, and detection rule engineering. To bridge this divide, the authors propose FORGE, a multi-agent system that introduces “exploit depth tiers” as a unifying framework for end-to-end vulnerability assessment. FORGE constructs targeted applications from CVE metadata, employs LLM-guided four-tier evaluation to iteratively refine exploits, and automatically generates Sigma and Snort detection rules from OpenTelemetry behavioral traces. Evaluated on 603 CVEs, FORGE achieves a 67.8% end-to-end success rate for L1+ exploits at a cost of $1.50 per CVE, produces L2+ detection rules that significantly improve normalized coverage (p=0.035), and yields Snort rules with zero false positives on benign traffic in 93.4% of cases.

0 citationsRead paper

Navigating the Sea of LLM Evaluation: Investigating Bias in Toxicity Benchmarks

May 11, 2026

Current benchmarks for evaluating toxicity in large language models exhibit underappreciated systematic biases that may lead to the deployment of unsafe models. This work systematically investigates how variations in task formulation—such as text completion versus summarization—input data domains, and evaluated models interact with multiple toxicity metrics. It reveals, for the first time, that both task type and data domain significantly influence toxicity scores. Experiments demonstrate that existing benchmarks are prone to misclassifying content as harmful when tasks are altered and show inconsistent performance across domains, highlighting their fragility and dependence on specific model-task configurations. These findings underscore the urgent need for more robust and reliable toxicity evaluation frameworks.

0 citationsRead paper