The security of Critical Infrastructures is increasingly threatened by the complexity of heterogeneous network devices. Although penetration testing remains the gold standard for verifying device security, the inherent limitation derived from the need for human operators to perform testing makes this activity resource-intensive and unscalable. To address these challenges, this paper proposes a fully automated four-stage framework for assessing the Node Compromise Risk. The proposed architecture operates as a progressive pipeline, starting from lightweight vulnerability mapping to resource-intensive verification phases based on Generative AI. At each stage, a dynamic metric updates the estimated risk, allowing the system to stop the analysis for safe nodes and allocate computational resources solely to high-risk targets. Generative AI is used in the final stage to dynamically adapt raw Proof-of-Concept exploits to the specific target configuration, bridging the gap between theoretical vulnerability and verified compromise. Experimental results conducted in a controlled environment demonstrate the framework’s ability to autonomously identify, filter, and exploit vulnerabilities, effectively reducing uncertainty about the security posture of a node.
Automatic Assessment of Node Compromise Risks in Critical Infrastructures via a Multi-stage Uncertainty Reduction Model Using LLMs / Artuso, D., Lax, G.. - 3020:(2026), pp. 422-436. (21st International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, IPMU 2026 ita 2026) [10.1007/978-3-032-28997-1_30].
Automatic Assessment of Node Compromise Risks in Critical Infrastructures via a Multi-stage Uncertainty Reduction Model Using LLMs
Lax G.
2026-01-01
Abstract
The security of Critical Infrastructures is increasingly threatened by the complexity of heterogeneous network devices. Although penetration testing remains the gold standard for verifying device security, the inherent limitation derived from the need for human operators to perform testing makes this activity resource-intensive and unscalable. To address these challenges, this paper proposes a fully automated four-stage framework for assessing the Node Compromise Risk. The proposed architecture operates as a progressive pipeline, starting from lightweight vulnerability mapping to resource-intensive verification phases based on Generative AI. At each stage, a dynamic metric updates the estimated risk, allowing the system to stop the analysis for safe nodes and allocate computational resources solely to high-risk targets. Generative AI is used in the final stage to dynamically adapt raw Proof-of-Concept exploits to the specific target configuration, bridging the gap between theoretical vulnerability and verified compromise. Experimental results conducted in a controlled environment demonstrate the framework’s ability to autonomously identify, filter, and exploit vulnerabilities, effectively reducing uncertainty about the security posture of a node.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


