PRACTICALLY LEVERAGING LLMS FOR MANUFACTURING CYBERSECURITY

Cybersecurity in manufacturing faces increasing threats and skilled personnel shortages. Large language models (LLMs), especially multi-modal variants, offer significant potential to rapidly parse complex data and identify vulnerabilities. This study explores the deployment of multi-modal LLMs in manufacturing cybersecurity, emphasizing their ability to bridge knowledge gaps and provide actionable insights. We evaluated offline and cloud-based models across two use cases, an analysis of a 100-page digital thread handbook used at a manufacturing facility, and vulnerability remediation in a manufacturing plant. The results highlight trade-offs between data privacy and model capability in understanding and prioritizing cybersecurity risks. Vision-based LLM limitations were evidenced through diagram analysis failures such as building layouts and network architecture diagrams, underscoring the need for human oversight and model transparency. Specialized cybersecurity embeddings showed promise for nuanced vulnerability analysis but still lack the ability to formally analyze multi-modal documents. Our findings emphasize current strengths, limitations, and pathways for using out-of-the-box LLMs and agentic artificial intelligence (AI) among industrial cybersecurity frameworks.