Gang Wang

The fast development of artificial intelligence (AI) presents new opportunities for both offensive and defensive use of AI in the security domain. In this talk, I will present several case studies to demonstrate (1) the emerging security threats introduced by AI-powered techniques, and (2) the challenges to applying AI for security defense. First, I will discuss new attacks enabled by AI models, using an example of face de-anonymization attacks. I will show how specially customized Diffusion models can be used to restore and re-identify Gaussian-blurred face images, compromising the intended privacy protection. Second, I will discuss the challenges and possible solutions to secure AI-assisted coding while defending against vulnerable/malicious code generation. I will describe our recent work to align Code LLMs to perform safety reasoning via rule learning to defend against the aforementioned threats. I will also share our experience of developing and testing this model as part of the Amazon Nova AI Challenge (2025). Finally, I will discuss general and open challenges in offensive and defensive use of AI for security.