Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

Glossary

Image scaling attack

A picture that looks harmless at full size but shows a prompt injection once the AI system scales it down to the size the model reads. Trail of Bits showed it against Gemini CLI, Vertex AI Studio and other Google systems in August 2025.

The user never sees the small image the model gets. The researchers' advice: do not downscale at all and limit upload sizes, or always show the user a preview of what the model actually sees, and never let text in an image start a sensitive tool call without confirmation.

Open the glossary in the Learning Center