Abstract

As Large Language Models integrate into Cyber–physical system like smart grids, their semantic vulnerabilities pose severe threats to physical infrastructure. Traditional digital domain alignment evaluations fail to capture these concrete physical risks. To address this, this paper proposes a transdomain dynamic quantitative evaluation framework driven by statistical mechanics and Monte Carlo simulations. Based on national grid standards, we construct a high fidelity adversarial dataset to test advanced structural transformation and logic driven coercion attacks. Furthermore, we introduce a Hazard Index (H) mapped to physical topologies and an entropy driven dynamic defense threshold (Ω). Through 10,000 concurrent simulations, our results reveal that traditional attack success metrics suffer from severe spurious positives. We find that high structural entropy injections drastically squeeze the attentional computing power of models, triggering a precipitous collapse of defense thresholds, whereas massive parameter scales provide critical attentional redundancy capacity to suppress physical penetration probabilities (P%). Consequently, static semantic filtering is fundamentally inadequate for industrial security. This study establishes a rigorous baseline for quantifying cascading physical risks, providing a solid theoretical foundation for the secure deployment of Artificial General Intelligence in future industrial internets.

IPC Classification

Keywords