Weitong Ruan

As AI systems become increasingly capable and autonomous, safety evaluation is emerging as a critical challenge. While current evaluations largely rely on benchmark performance and empirical red-teaming, these approaches provide limited guarantees about real-world risk.

As organizations deploy generative AI at scale, watermarking has become the critical trust layer for establishing content provenance, meeting emerging regulatory requirements, and protecting brand integrity. But what if an adversary could stamp a fake watermark onto any image and frame your service?