FlowHijack: A Dynamics-Aware Backdoor Attack on Flow-Matching Vision-Language-Action Models

Xinyuan An, Tao Luo, Gengyun Peng, Yaobing Wang, Kui Ren, Dongxia Wang

• Vision-Language-Action models relying on flow-matching vector-field dynamics suffer from unexamined security risks. We propose FlowHijack, the first backdoor attack targeting such continuous policies. It stealthily manipulates action generation while maintaining benign performance and kinematic normality, exposing critical vulnerabilities in embodied models and demanding dedicated defenses.

Watch Out Your Industrial Copilots: Stealthy Backdoor Attack Against LLM-Based PLC Code Generation

Xinyuan An, Xiaoxia Liu, Dongxia Wang, Zhanhang Xiong, Wenhai Wang

• LLM-based PLC code generation introduces unexamined security risks to industrial control systems. We propose STBack, the first stealthy backdoor attack targeting such industrial copilots. It leverages semantic-integrated triggers to stealthily generate compromised PLC code that bypasses formal verification while maintaining semantic normality , exposing critical physical vulnerabilities and demanding dedicated defenses.

CrossPL: Systematic Evaluation of Large Language Models for Cross Programming Language Interoperating Code Generation

Zhanhang Xiong, Dongxia Wang, Yuekang Li, Xinyuan An, Wenhai Wang