Honey, I shrunk the image and now I'm pwned

Trail of Bits demonstrated an image-scaling prompt-injection technique that embeds hidden instructions in large images which become visible to AI models when images are downscaled; they proved data-exfiltration against Google Gemini CLI and related interfaces, released a crafting tool (Anamorpher), and recommended avoiding downscaling or showing users what the model actually sees, while Google says the attack requires non-default configuration to work.