<p>Due to the temporal characteristics of industrial data, deep time series models are commonly employed in industrial anomaly detection systems. However, such models exhibit inherent security vulnerabilities when exposed to adversarial examples, posing significant threats to the reliability of industrial systems. Existing adversarial example attack methods are primarily designed for image-based tasks and typically induce noticeable perturbations. As industrial data exhibits stable distributions and structured characteristics, such perturbations are often flagged and eliminated by standard mechanisms like bad data detection (BDD) or state estimation (SE), rendering these attacks ineffective in industrial scenarios. Nonetheless, industrial anomaly detection models are not inherently robust, highlighting the need for adversarial attacks specifically tailored to industrial temporal and structural characteristics. This paper presents a novel adversarial attack method, Variable Perturbation Gradient Method (VPGM), tailored for industrial time series anomaly detection, with the Secure Water Treatment (SWaT) system serving as a representative case study. VPGM generates effective adversarial samples by adaptively adjusting perturbation magnitudes and ranking data points based on their temporal importance. Experimental results demonstrate that VPGM successfully compromises industrial anomaly detection systems, inducing both false alarms and missing alarms. Finally, we propose a universal bypass defense strategy that effectively defends VPGM attack and other previous attacks without halting industrial operations or requiring reconfiguration of the industrial anomaly detection systems.</p>