Get the latest tech news
Weaponizing image scaling against production AI systems
In this blog post, we’ll detail how attackers can exploit image scaling on Gemini CLI, Vertex AI Studio, Gemini’s web and API interfaces, Google Assistant, Genspark, and other production AI systems. We’ll also explain how to mitigate and defend against these attacks, and we’ll introduce Anamorpher, our open-source tool that lets you explore and generate these crafted images.
Furthermore, these algorithms are implemented differently across libraries (e.g., Pillow, PyTorch, OpenCV, TensorFlow), with varying anti-aliasing, alignment, and kernel phases (in addition to distinct bugs that historically have plagued model performance). Anamorpher provides users with the ability to visualize and craft image scaling attacks against specific algorithms and implementations through a front-end interface and Python API. It would also be useful to explore more effective fingerprinting approaches, semantic prompt injection, factorized diffusion, polyglots, and additional artifact exploitation, especially any typically chained with upscaling (such as dithering).
Or read this on Hacker News