Brief

A new research paper explores the vulnerabilities of large language models (LLMs) when applied to cyber threat intelligence (CTI). The study identifies three specific cognitive failures in LLMs within CTI workflows: spurious correlations from metadata, contradictory knowledge from conflicting sources, and limited generalization to new threats. Researchers developed a human-in-the-loop framework to label these failures and demonstrated that targeted defenses can significantly reduce error rates, offering a path toward more resilient CTI agents. AI

Researchers have developed a novel framework called Adversarial Network Imagination to proactively identify and simulate potential failures in telecommunication networks. This system utilizes a Causal Large Language Model (LLM) to generate realistic failure scenarios, which are then tested within a digital twin of the network. The goal is to move network management from a reactive approach to one focused on anticipatory resilience by evaluating mitigation strategies before actual disruptions occur. AI

Researchers have developed SG-LLM, a novel method for extracting patient data from electronic health records to create digital twins. This approach grounds LLM extraction with schema constraints and a validation loop for repair, improving the accuracy and validity of the generated FHIR bundles. An experiment on clinical utility demonstrated that classifiers trained on SG-LLM-generated data performed comparably to those trained on expert-curated data, suggesting its effectiveness in real-world healthcare applications. AI